rmbr

Give your agent memory and knowledge. One file, three lines, no server, no API key.

rmbr ("remember", vowels deleted) is an embedded, local-first memory + retrieval engine for AI agents and LLM apps — what SQLite is to Postgres, rmbr aims to be to hosted memory services.

v0.2.7. pip install rmbr gets you a working library: Memory, Index, Policy, MCP support, an optional HTTP server, PDF/DOCX ingestion, and framework adapters for LangChain/LlamaIndex/LangGraph/mem0 (all below), all implemented and tested — see docs/PLAN.md and docs/ARCHITECTURE.md for the design.

Start with three lines, then reach for exactly as much more as you need — nothing below is required to use the part above it:

• Memory — durable, searchable notes your agent chooses to keep, namespaced per agent
• Index — hybrid (keyword + semantic) search over your own docs, for RAG
• Policy — deny-by-default access control, so one agent can't read another's memory unless you explicitly allow it
• Framework adapters — LangChain and LlamaIndex retrievers, a real LangGraph BaseStore, a mem0-API-compatible drop-in, raw OpenAI/Anthropic tool-calling export
• An MCP server, for any MCP client (Claude Desktop, Claude Code, Cursor, ...) — optional
• An HTTP server, for serverless functions or anything that'd rather curl it than hold a connection open — also optional

If you only ever use the first three, that's not a "basic" use of rmbr — that is rmbr for most people. The server modes exist for the specific cases they solve, not because you're expected to grow into them.

Contents: Why · Quickstart · Multi-agent isolation · MCP support · HTTP support · Alternatives · Performance · Roadmap

Why

Agents can already "remember" things across restarts — a CLAUDE.md, a system prompt, a JSON file on disk. That's not new, and rmbr isn't claiming otherwise.

What breaks is what happens as that file grows. Every fact in a static context file costs tokens on every single call, whether it's relevant to the current task or not — so it either stays small (a few dozen hand-curated notes) or turns into noise nobody's cheaply reading anymore. There's no ranking: the agent gets the whole file, or nothing, never just the 5 facts that actually matter for this turn. A static file gets more expensive and less useful the more the agent learns; a searchable memory gets more useful and stays the same cost per call. mem.recall(query) returns the k most relevant memories out of however many thousand you've accumulated — that's the actual gap between "an agent that can write to a file" and "an agent with memory."

The other place people get burned: rolling this yourself. Chunk text, embed it, throw it in a vector store — that's a legitimately easy weekend project (this one started that way too). What's easy to get wrong in that weekend project: real hybrid search (most ship vector-only or keyword-only and never notice), an embedding cache (so you're not re-embedding — and re-paying for — the same text on every call), and, if there's more than one agent involved, safe isolation between them. Most hand-rolled or framework-provided multi-agent memory either shares one blob every agent can read and write, or scopes access via a namespace/user_id parameter the calling model itself supplies — which a prompt injection can simply ask to change. rmbr's MCP tools don't expose that parameter at all; there's no field for an injected instruction to fill in.

So: rmbr exists for the gap between "stuff it in a system prompt" (doesn't scale past a few KB) and "stand up real infrastructure" (Docker, a graph database, a hosted API key) — search-quality, safely-isolated memory, as a dependency, not a service.

Concretely, rmbr gives you:

• One file. Your agent's entire memory and knowledge base is a single .db file — git commit it, diff it, roll it back, hand it to a teammate, attach it to a bug report, or check a known-good state into a test fixture for deterministic CI. No hosted memory service lets you do any of that.
• Three lines. pip install rmbr, import, remember. No account, no config, no service.
• No added infrastructure. Your agent already needs a network connection and an API key for its LLM calls — rmbr doesn't add a second one just for memory. mem0 defaults to a hosted LLM+embedding API, Zep needs Docker+Neo4j+an LLM key, Letta needs a server+Postgres — all on top of whatever you're already paying for the model itself. rmbr's own memory/retrieval path makes zero network calls by default: one less vendor, one less key to leak, one less service whose outage takes your agent's memory down with it. (It also means rmbr keeps working with a fully local LLM — Ollama, llama.cpp — for genuinely offline or air-gapped use; most people won't need that, but it's there.)
• No proprietary format. rmbr never calls an LLM itself — recall()/search() return plain strings, floats, and dicts (hit.text, hit.score, hit.metadata). Nothing to parse, no vendor SDK required to consume it — see Using results with an LLM below for how that plugs into Claude, GPT, or Gemini identically.
• Namespace-pinned multi-agent access. Policy is deny-by-default; MCP tools expose no namespace parameter to override — safe by construction, not by convention.

Quickstart

from rmbr import Memory

mem = Memory("agents.db", namespace="assistant")
mem.remember("user prefers dark mode and short answers")
mem.recall("user preferences")

Three lines — that's the whole API for the common case. Everything below is opt-in and lives in its own section, so you only read what you actually need. Library-only by design — no CLI to learn. (python -m rmbr exists solely so MCP clients can launch the server; see MCP support below.)

agents.db doesn't need to exist first. There's no rmbr init, no template to download, nothing to provision — Memory(path, ...) (and Index(path)) create the file the moment you call them on a path that doesn't exist yet, with the right schema already in place. The one thing that does need to exist is the directory the path lives in (same as opening any file for writing) — Memory("agents.db", ...) works from wherever you run it; Memory("some/deep/agents.db", ...) needs some/ to already be there.

Indexing documents (RAG)

from rmbr import Index

idx = Index("agents.db")
idx.add_files("docs/")                     # .py, .md, and plain text each get an appropriate splitter automatically
hits = idx.search("how do I deploy?", k=5)
hits[0].text, hits[0].score, hits.timings  # per-stage latency, always visible

Index and Memory share the same .db file — open both against the same path if your agent needs a knowledge base and a memory. add_files()/add_texts() return an IngestResult: a plain list of document ids with a .timings breakdown attached (chunk_ms/embed_ms/store_ms/ann_ms/docs_per_second) — the same transparency hits.timings gives you for search, applied to ingestion, so you can see for yourself that embedding dominates the cost rather than take our word for it.

Using results with an LLM

rmbr never calls a model — search()/recall() hand you back plain text and a score, and you decide what to do with it. The standard pattern (classic RAG: retrieve, then inject the retrieved text into the prompt) with Claude:

from anthropic import Anthropic
from rmbr import Index

idx = Index("agents.db")
idx.add_files("docs/")

client = Anthropic()  # reads ANTHROPIC_API_KEY from the environment

def answer(question: str) -> str:
    hits = idx.search(question, k=5)
    context = "\n\n".join(f"<document>{hit.text}</document>" for hit in hits)
    response = client.messages.create(
        model="claude-sonnet-5",
        max_tokens=1024,
        # Context before the question, not after — Anthropic's own prompting
        # docs measure this ordering as meaningfully better for long-context
        # RAG, though it isn't required for correctness.
        messages=[{"role": "user", "content": f"{context}\n\nUsing the documents above, answer: {question}"}],
    )
    return response.content[0].text

answer("how do I deploy?")

This isn't Claude-specific. hit.text is a plain Python string with no wrapper, no provider object, nothing rmbr-proprietary — the exact same context string above drops verbatim into OpenAI's messages array (client.chat.completions.create(model=..., messages=[...])) or Gemini's contents. Every mainstream chat-completion API takes the same fundamental shape (a list of role-tagged text messages), which is why "retrieve text, put it in the prompt" — the only integration contract rmbr makes — works identically across providers. Swap the SDK call, nothing else changes.

Want the embedding itself to come from a hosted provider instead of the local default? Memory("agents.db", namespace="assistant", embedder=OpenAIEmbedder()) (pip install rmbr[openai]) — VoyageEmbedder/pip install rmbr[voyage] and CohereEmbedder/pip install rmbr[cohere] are also available, all three behind the exact same Embedder protocol, same rest of the API.

Keeping memory accurate over time

remember() inserting forever is fine for a while, then it isn't: near-duplicate notes pile up, and nothing ever expires. rmbr doesn't have an LLM to judge "is this the same fact" the way mem0's extraction loop does — everything below is deterministic vector-similarity/time-based engineering instead, opt-in because a false-positive match is a worse failure than a duplicate:

# Update-in-place instead of appending, above a cosine-similarity threshold.
# Off by default — no LLM here to judge intent, so keep it conservative (0.92-0.95).
mem = Memory("agents.db", namespace="assistant", dedupe_threshold=0.93)
mem.remember("user prefers dark mode")       # inserts
mem.remember("user really prefers dark mode")  # updates the same row if similarity clears the bar

# Bound growth automatically (evicts the oldest beyond the cap on every remember()),
# or prune on your own schedule:
mem = Memory("agents.db", namespace="assistant", max_memories=5000)
mem.forget_older_than(60 * 60 * 24 * 30)     # delete anything older than 30 days

# max_memories eviction is pure recency by default, which is a real risk once
# you rely on it - a trivial fact from an hour ago would otherwise outlive a
# critical one from last week for no reason but timestamp. Exempt specific
# memories from it:
mem.remember("the customer's account was permanently deactivated", pinned=True)

Loading many items into an already-large namespace (an org's internal doc set, a backfill of historical memories) is a different situation than a single remember() mid-conversation — batch it:

# Without this, every single remember()/add_text() call re-serializes the
# *entire* vector index (usearch has no incremental on-disk save) - fine
# at hundreds-to-low-thousands per namespace, real cost once a namespace
# has tens of thousands of items and you're adding many more sequentially.
with mem.bulk():
    for fact in many_facts:
        mem.remember(fact)
# SQL rows are still durable immediately inside the block; only the vector
# index's persistence is deferred to one write when the block exits - a
# crash mid-block loses whatever hadn't been flushed yet, a real tradeoff
# you're opting into, not a silent one. See Performance below for numbers.

Index has the same .bulk().

Check on a namespace's memory without hand-writing SQL:

mem.stats()                       # {"assistant": {"count": 412, "oldest": "...", "newest": "..."}}
mem.stats(namespaces="*")         # same, broken down per namespace this policy can read
mem.integrity_check()             # [] if healthy; otherwise, what's wrong and which ids

Index has the same two methods, reporting documents/chunks counts instead.

Precision knobs for search

search()/recall() default to plain hybrid ranking, but three things are available when relevance quality matters more than the default:

# Richer where= filtering: equality by default, $eq/$ne/$gt/$gte/$lt/$lte/$in/$nin as operators.
idx.search("deploy", where={"updated_at": {"$gt": "2026-01-01"}})

# A real confidence gate — filters on the raw cosine similarity (hit.vector_score),
# not hit.score itself, which is an RRF rank-sum with no fixed scale to threshold on.
idx.search("deploy", min_similarity=0.6)

# Recency-weighted ranking: a fresher memory/chunk can outrank an equally
# relevant older one. recency_weight=0.0 (off) by default. A chunk's "created"
# time is its document's ingestion time (add_text()/add_files()).
mem.recall("user preferences", recency_weight=0.05, recency_half_life_seconds=7 * 86400)
idx.search("deploy", recency_weight=0.05)

# A local cross-encoder re-scores the candidate pool for higher precision at
# extra latency — same fastembed dependency already installed, no new network
# call, no API key. hit.score becomes the cross-encoder's score when this is on.
idx.search("deploy", rerank=True)

Conversation memory

The most common real agent shape is a chat loop that should remember across turns. remember_turn() is a thin convenience over remember() for exactly that — role/session_id land in metadata rather than getting baked into the stored text, so semantic search isn't polluted by a "user: " prefix and you can filter or replay by either:

mem.remember_turn("user", "I prefer dark mode")
mem.remember_turn("assistant", "Got it, dark mode from now on", session_id="conv-42")

mem.recall("dark mode", where={"role": "user"})       # who said it
mem.list(where={"session_id": "conv-42"})              # replay one conversation, in order (list(), not recall() — no query needed)

Wiring into an existing agent loop or framework

Three ways to plug rmbr into whatever's already running your agent, without going through MCP:

# Raw OpenAI/Anthropic tool-calling — one line to get a ready-made tool
# definition plus a callable, in either API's shape:
tool = idx.as_tool()
response = client.messages.create(..., tools=[tool.to_anthropic()])
result = tool.call(**tool_use_block.input)          # dispatches to idx.search()

recall_tool, remember_tool = mem.as_tools()          # or as_tools(read_only=True) for recall only

# LangChain — wraps Index as a real BaseRetriever, drops into any chain
# (pip install langchain-core, or whatever LangChain distribution you're on):
retriever = idx.as_langchain_retriever(k=5)
retriever.invoke("how do I deploy?")                 # -> list[Document]

# LlamaIndex — same idea (pip install llama-index-core):
retriever = idx.as_llamaindex_retriever(k=5)
retriever.retrieve("how do I deploy?")                # -> list[NodeWithScore]

# LangGraph — a real BaseStore, drops into StateGraph(...).compile(store=...)
# (pip install langgraph-checkpoint):
from rmbr.integrations.langgraph import as_store
store = as_store("agents.db")
store.put(("memories", "user-42"), "pref-1", {"text": "user prefers dark mode"})
store.search(("memories", "user-42"), query="dark mode")   # -> list[SearchItem]

Both retriever adapters accept the same search() keyword arguments (where=, min_similarity=, rerank=, ...) and have async equivalents (retriever.ainvoke(...) / retriever.aretrieve(...), backed by Index.asearch()). Neither langchain-core nor llama-index-core is a required rmbr dependency — each adapter imports its target framework lazily, only when you actually call as_langchain_retriever()/as_llamaindex_retriever().

as_store() maps a LangGraph namespace tuple to one rmbr namespace (joined by .), and a LangGraph key to metadata["_lg_key"] — see rmbr/integrations/langgraph.py's module docstring for the exact mapping and what's deliberately not supported (per-item TTL, field-path-selective indexing). langgraph-checkpoint isn't a required rmbr dependency either.

Coming from mem0

rmbr.integrations.mem0_compat.Memory matches mem0 OSS's local Memory class call-for-call (add()/search()/get_all()/get()/update()/delete()/delete_all(), same argument names, same {"results": [...]} / {"message": "..."} return shapes) so most of an existing mem0 integration ports by changing the import and the constructor call:

from rmbr.integrations.mem0_compat import Memory   # was: from mem0 import Memory

m = Memory("agents.db")                              # was: Memory()  (rmbr writes to a file you name)
m.add("user prefers dark mode", user_id="alex", infer=False)
m.search("dark mode", filters={"user_id": "alex"})

This isn't a wrapper around mem0 — no mem0ai dependency, not even optional. One behavior is a deliberate hard no rather than a silent difference: mem0's real default infer=True has an LLM read your messages and decide what to keep; rmbr never calls an LLM, so add(..., infer=True) (or leaving infer unset — mem0's own default) raises NotImplementedError naming exactly what's not happening, rather than quietly storing raw text under an argument that claimed something smarter was going on. Pass infer=False to store messages as-is. See the module docstring for the full list of what's matched, what's translated (filters={"key": {"gt": 10}} -> rmbr's where=), and what's unsupported (mem0's AND/OR/NOT filter combinators, history(), vision messages).

as_tool()/as_tools()'s exported schema isn't limited to query/k — a calling model can also pass where/min_similarity/rerank on any given call (all optional, so a model that doesn't know about them behaves exactly as before):

tool.call(query="how do I deploy?", where={"tier": "public"}, min_similarity=0.6, rerank=True)

Every built-in schema sets additionalProperties: false, and tool.call() validates arguments against it before dispatching — a model that hallucinates an argument (smaller/faster models do this more than you'd hope) gets back a ToolCallError naming the actual problem, safe to feed straight back as a tool-result error, instead of a bare Python TypeError taking down your process. For providers that support it, to_anthropic(strict=True) / to_openai(strict=True) asks the provider itself to reject a malformed call before it's even dispatched — a complement to, not a replacement for, call()'s own validation, since not every provider enforces strict as tightly as it's documented to.

Restricting access between agents

from rmbr import Memory, Policy

policy = Policy()
policy.allow("supervisor", read="*")  # supervisor can read every namespace

mem = Memory("agents.db", namespace="coder", policy=policy)

Deny-by-default: coder can only read/write its own namespace unless explicitly granted. See Multi-agent isolation below for the full model, the security reasoning, and a diagram of a real team topology.

Async, for web backends and concurrent agents

Every read and write has an a-prefixed async twin — aremember/arecall/aforget on Memory, aadd_text/aadd_texts/aadd_files/asearch on Index — for async def route handlers (FastAPI, Starlette, aiohttp) where a blocking call stalls every other request on the same event loop:

from fastapi import FastAPI
from rmbr import Memory

app = FastAPI()
mem = Memory("agents.db", namespace="assistant")

@app.post("/chat")
async def chat(message: str):
    context = await mem.arecall(message, k=5)
    await mem.aremember(f"user said: {message}")
    return {"context": [hit.text for hit in context]}

Or fan a supervisor out across several granted namespaces concurrently instead of one at a time:

import asyncio

coder_notes, researcher_notes = await asyncio.gather(
    supervisor.arecall("release blockers", namespaces="coder"),
    supervisor.arecall("release blockers", namespaces="researcher"),
)

One honestly-stated limitation: async calls on the same Memory/Index instance are serialized behind an internal lock, reads included. That's deliberate — the vector index (usearch) isn't documented as safe for concurrent mutation from multiple threads, and a corrupted index is a far worse failure than giving up some read concurrency. Open separate instances against the same file for true parallelism; SQLite's WAL mode supports that fine.

Serving memory over MCP

from rmbr import serve_mcp

serve_mcp("agents.db", namespace="coder", read_only=True)

See MCP support below for what this exposes and how to actually connect a client to it.

Serving memory over HTTP (optional)

from rmbr import serve_http

serve_http("agents.db", namespace="coder", read_only=True, token="a-shared-secret")

For callers that can't be an MCP client and can't import rmbr either — a serverless function, a process on another machine, anything that would rather curl a URL than hold a connection open. You don't need this to use rmbr — it's an alternative front door onto the same Memory/Index, not a requirement layered on top of them. See HTTP support below for the full endpoint list, the auth story, and why it costs zero new dependencies.

Contributing / running from source

git clone https://github.com/SRock44/rmbr.git
cd rmbr
python -m venv .venv && source .venv/bin/activate   # .venv\Scripts\activate on Windows
pip install --only-binary :all: -e .
pytest tests/    # 284 tests, no network or API key required

The default embedder (fastembed, a local ONNX model) downloads its model weights on first use. Every test in tests/ instead uses rmbr.embed.FakeEmbedder — a deterministic, dependency-free embedder — so the suite runs fully offline; you can inject the same FakeEmbedder into your own tests via Memory(..., embedder=FakeEmbedder()) / Index(..., embedder=FakeEmbedder()).

Multi-agent isolation, honestly stated

• Namespaces keep agents' memories separate and are enforced on every call — but they are organizational, not cryptographic. Any code with access to the file can open the file. That's true of every embedded database; we say it out loud.
• Hard isolation = separate .db files per trust boundary, plus OS file permissions.
• MCP serving is namespace-pinned: the exposed tools have no namespace parameter, so an external agent structurally cannot query outside its lane — unlike every other MCP memory server we looked at, where the scope is a parameter the calling model supplies (and could be talked into changing).

A concrete team topology — one supervisor with a broad grant, two specialists that can't see each other, one external MCP client pinned to a single lane, all in the same agents.db file:

flowchart TB
    subgraph db["agents.db — one SQLite file"]
        direction LR
        supNS[("supervisor<br/>namespace")]
        coderNS[("coder<br/>namespace")]
        researchNS[("researcher<br/>namespace")]
    end

    supervisor["Supervisor agent<br/>policy.allow('supervisor', read='*')"] ==>|read + write| supNS
    supervisor -.->|read, explicitly granted| coderNS
    supervisor -.->|read, explicitly granted| researchNS

    coder["Coder agent<br/>Memory(path, namespace='coder')"] ==>|read + write| coderNS
    researcher["Researcher agent<br/>Memory(path, namespace='researcher')"] ==>|read + write| researchNS

    external["External MCP client<br/>(Claude Code, Cursor, ...)"] -->|"serve_mcp(path, namespace='coder')"| coderNS

The coder and researcher namespaces have no path between them on this diagram — that's the point, not an omission. Nothing needed to be configured to deny that access; only the supervisor's grant (read="*") is explicit. The external MCP client's tool schema has no namespace argument at all, so it structurally cannot ask for anything outside coder, no matter what a document it's summarizing tells it to try.

See examples/multi_agent_support/ for this pattern as a runnable end-to-end demo — three Claude-powered agents (two isolated specialists + a supervisor) sharing one .db file, including a live PermissionError when isolation is tested directly against the API.

MCP support

MCP (Model Context Protocol) is an open, model-agnostic protocol for connecting AI applications — Claude Desktop, Claude Code, Cursor, and a growing list of others — to external tools and data sources through one standard interface, instead of every app inventing its own plugin format. rmbr speaks MCP so any MCP-capable client can search and remember through your .db file directly, without you writing a server yourself.

What serve_mcp() exposes

from rmbr import serve_mcp

serve_mcp("agents.db", namespace="coder")                  # read + write
serve_mcp("agents.db", namespace="coder", read_only=True)  # read only

Three tools, all pinned to whatever namespace you pass at startup (see Multi-agent isolation above for why there's no namespace parameter for a client to override):

• search(query, k=5) — hybrid search over documents added via Index
• recall(query, k=5) — search over notes saved via Memory
• remember(text, pinned=False) — save a new memory; pinned=True exempts it from max_memories eviction. Not present in the tool list at all — not just permission-denied — when read_only=True.

Each result includes bm25_score/vector_score (the raw signals behind score) alongside text/metadata — useful if the calling agent wants to weight or filter results by confidence rather than trust every hit equally. min_similarity, recency_weight, and rerank (see Precision knobs for search above) aren't exposed as MCP tool parameters yet — the tool schemas stay minimal on purpose; configure them at serve_mcp()'s call site via a custom Index/Memory if you need them server-side.

Also exposed: an MCP resource template, rmbr://examples/{pattern} (plus rmbr://examples listing the valid pattern values), serving short, runnable code snippets for common usage patterns — basic-memory, document-search, multi-agent-policy, conversation-memory, tool-calling, memory-hygiene. Any MCP client that can browse resources (not just call tools) can pull these up directly, without leaving the session or going to GitHub.

Connecting a client

serve_mcp() blocks on stdio; it's meant to be launched as a subprocess by an MCP client, not called from inside your own long-running app. python -m rmbr is the launch shim for exactly that (the package also installs a rmbr console script pointing at the same thing, so uvx rmbr works without a local install):

python -m rmbr agents.db --namespace coder --read-only
# or, via the installed console script / uvx:
rmbr agents.db --namespace coder --read-only

For Claude Desktop or Claude Code, add it to your MCP config (Claude Desktop's claude_desktop_config.json, or a project's .mcp.json):

{
  "mcpServers": {
    "rmbr-coder": {
      "command": "uvx",
      "args": ["rmbr", "/absolute/path/to/agents.db", "--namespace", "coder", "--read-only"]
    }
  }
}

Restart the client and its tool list picks up search/recall (and remember, unless read-only) scoped to that one namespace. The rest of the file — every other agent's memory — isn't reachable through this connection; there's no parameter that would let it be.

HTTP support

This entire section is optional. Everything above it — Memory, Index, Policy, MCP — works with no HTTP server anywhere in the picture, and that's how most rmbr users actually run it: import the library, call a few methods, done. Nothing about serve_http() existing changes that; it's not a more "grown-up" way to use rmbr, it's a different front door for a specific situation the ones above don't cover.

That situation: a caller that's in a different process, on a different machine, or can't hold a connection open the way an MCP client does. MCP expects a client to launch serve_mcp() as a subprocess it owns via stdio — a serverless function that spins up per-request can't do that. And if rmbr's .db file lives somewhere your caller's process doesn't (a different container, a different machine entirely), import rmbr isn't an option either. What is always an option: an HTTP request. That's the entire reason serve_http() exists — nothing more.

If neither of those describes what you're building, you can stop reading here — rmbr isn't nudging you toward running a server.

Starting it

from rmbr import serve_http

serve_http("agents.db", namespace="coder", read_only=True, token="a-shared-secret")

Blocks until stopped — same as serve_mcp(), this is meant to be your process's entire job, not something called from inside an app that's also doing other work. Binds to 127.0.0.1 by default; pass host="0.0.0.0" only once you've actually decided this should be reachable from outside this machine.

Zero new dependencies. Starlette and uvicorn aren't something rmbr added for this — mcp (already a hard rmbr dependency, for its own HTTP transport) pulls both in already. Turning on serve_http() doesn't grow your dependency tree by a single package.

What it exposes

Namespace-pinned, the same principle as serve_mcp(): no request body or query string anywhere in this API has a namespace field, so a caller structurally cannot reach outside the one namespace this server was started for — see Multi-agent isolation above for why that matters more than it might sound like it does.

add_files() isn't on this list on purpose — it reads from this process's local filesystem, which is meaningless to a caller on the other end of an HTTP request. Send the text itself to POST /documents instead. Every write route returns 405 when the server was started with read_only=True, same semantics as serve_mcp()'s read_only hiding the remember tool entirely.

Talking to it needs nothing but curl:

curl -X POST http://127.0.0.1:8000/memories \
  -H "Authorization: Bearer a-shared-secret" \
  -H "Content-Type: application/json" \
  -d '{"text": "user prefers dark mode"}'
# {"id": 1}

curl -X POST http://127.0.0.1:8000/memories/search \
  -H "Authorization: Bearer a-shared-secret" \
  -H "Content-Type: application/json" \
  -d '{"query": "dark mode"}'
# {"results": [{"id": 1, "text": "user prefers dark mode", "score": ..., ...}], "timings": {...}}

Auth is opt-in, not automatic

Pass token= (or set the RMBR_TOKEN environment variable) and every route except /health requires Authorization: Bearer <token>; leave both unset and there is no auth at all. That default matches rmbr's posture everywhere else — you own the network boundary, rmbr doesn't assume one for you — but it's worth being deliberate rather than just accepting the default: if you're binding to anything other than 127.0.0.1, set a token.

Composing it into something bigger

serve_http() is a thin, blocking convenience wrapper around build_app(), which hands back a plain Starlette application — nothing rmbr-proprietary about it:

from rmbr.server import build_app

app = build_app("agents.db", namespace="coder")
# it's just an ASGI app from here: mount it inside a larger Starlette/FastAPI
# app, wrap it in your own middleware (CORS isn't included - add
# starlette.middleware.cors.CORSMiddleware yourself if you need it), or hand
# it to a different ASGI server entirely instead of calling serve_http().

Full design notes (why namespace-pinned, what the auth middleware does, what deliberately isn't supported) live in rmbr/server.py's module docstring.

Alternatives

Not "competitors" — genuinely different tools for genuinely different jobs. Here's where each one actually fits, including where rmbr isn't the right choice.

If you're evaluating a memory service (mem0, Zep/Graphiti, Letta): all three are excellent at LLM-mediated memory intelligence — extracting facts from conversation, resolving contradictions, consolidating duplicates. rmbr deliberately does none of that; it never calls an LLM, full stop. That's a real capability gap, not spin — but it's also why rmbr has no API key requirement, no extra LLM cost or latency on every remember(), and no risk of a consolidation model quietly rewriting what you actually said. You get the primitives (remember/recall/forget, namespace policy); you decide what, if anything, sits on top.

(GitHub stars as of this writing, for scale: mem0 ~62k, Graphiti ~29k, Letta ~24k. This is a much larger, faster-moving category than rmbr is part of — worth knowing going in.)

If you're evaluating a vector database (Chroma, LanceDB, pgvector, Pinecone, ...): these are real peers on "embedded, no API key" — Chroma and LanceDB in particular are just as zero-server as rmbr. The difference is what's built on top of the vector index: with a raw vector database you're still building the memory API, the namespace/access-control layer, the hybrid BM25+vector fusion, the embedding cache, and an MCP server yourself. rmbr ships all of that already assembled, specifically for the agent-memory shape of problem.

Where they legitimately win: raw bulk-ingestion throughput at large scale. If you're indexing millions of documents for a dedicated search product, use a purpose-built vector database — that's their job, not rmbr's. rmbr is tuned for what an agent's own memory and knowledge base actually looks like (its own history, a knowledge base in the hundreds-to-low-thousands of chunks), where single-call latency, not bulk-loading speed, is what you actually pay for on every turn. See Performance below for the honest numbers on both.

Performance

This README will never contain a performance number that isn't produced by a script in bench/ — reproducible by anyone, on disclosed hardware, methodology included.

The number that matters for rmbr's actual usage pattern — an agent calling remember()/search() one at a time mid-reasoning-loop, not bulk-loading a corpus — is single-call latency with the real default embedder, not bulk throughput. That's what's below, run on the project's pinned Ubuntu benchmark machine (Intel Core Ultra 9 285K, 4 cores isolated via taskset -c 0-3, Ubuntu 24.04.4 LTS, Python 3.12.3), median of 3 runs, 100 samples/run:

Read that third row carefully: ~85-90% of a plain search call's cost is the embedding model, not rmbr. rmbr's own storage/retrieval overhead is sub-millisecond. And all of this is imperceptible next to the LLM call that will follow it in any real agent loop — which was rmbr's founding thesis about where RAG latency actually lives (see docs/PLAN.md).

The last two rows are what v0.2's rerank=True and recency_weight actually cost on top of a plain search call. rerank=True is real, measured cost — a local cross-encoder pass over the candidate pool — because it's doing genuine additional inference, not a free re-sort; its p95/p99 run noticeably higher than its p50 because the reranker model lazy-loads (and, on a cold cache, downloads) on an index's first rerank=True call, not at import time — use it when result quality matters more than shaving milliseconds, not on every call by default. recency_weight is effectively free (same latency as a plain search, within noise), since it's pure-Python exponential decay math over chunks already fetched, no extra model call. Reproduce: python bench/latency.py --n-calls 100 --n-queries 100 --corpus-size 500; raw output for all 3 runs is in bench/pinned/.

Bulk-ingest throughput, for full transparency (not a claim we're leading with): rmbr batches every write in add_texts()/add_files() into one SQLite transaction, one embedder call, and one ANN-index insert for the whole batch, rather than once per document — a real, measured ~2,966 docs/s (hybrid, default; median of 3 seeds) on a 5,000-doc synthetic corpus. Note what didn't move much: batching the embed call barely helped in this specific benchmark, because it feeds every engine identical precomputed vectors (a near-free dict lookup) specifically to isolate storage/ANN performance — a real embedder (ONNX inference, or an API call) has real fixed per-call overhead that batching actually amortizes, so bench/latency.py's numbers above are the more representative ones for real-world embedding cost.

Against the two purpose-built vector databases, rmbr is still slower at pure bulk loading — a fundamentally different job than what rmbr is built for: Chroma ingests ~2.6x faster (~7,775 docs/s median) and LanceDB ~35-80x faster (~104,000-236,000 docs/s, wide variance across runs), because it's one Arrow batch write with zero per-row relational bookkeeping. Against mem0 — the closer peer, since it's an actual memory abstraction, not a raw vector store — the result flips: rmbr ingests ~7.4x faster (~2,966 vs ~401 docs/s median), reflecting mem0's real per-row cost (a SQLite history/audit-log write plus a BM25 sparse-vector encode alongside the dense one, on every insert, left on for this benchmark since that's mem0's real default — see Coming from mem0 above for why rmbr does neither by default). What rmbr does hold its own on across all three: recall@5 (0.949) is close behind mem0's hybrid search (0.998) and LanceDB's exact search (1.000), and clearly ahead of Chroma's vector-only search (0.797). Full numbers, all 3 seeds (now including mem0), in bench/pinned/ and reproducible via pip install -e ".[bench]" && python bench/run.py. We're disclosing this, not hiding it: if bulk document loading at scale is your actual workload, see Alternatives above — that's not what rmbr optimizes for.

Scale: what happens once a namespace holds tens of thousands of items

usearch (the vector index) has no incremental on-disk save — every remember()/add_text() call re-serializes and rewrites the entire vector index, every time. At rmbr's normal scale (hundreds to low-thousands per namespace) that's negligible. Once a namespace grows into the tens of thousands, many sequential writes each pay to reserialize everything that came before — real, measured, and now fixed with Memory.bulk()/Index.bulk() (see Keeping memory accurate over time above for usage). Cost of 50 sequential remember() calls into an already-populated namespace, with vs. without .bulk():

Without .bulk(), per-write cost climbs linearly with namespace size — the signature of the O(n) reserialize happening on every call. With it, per-write cost barely grows (0.81ms → 6.82ms across a 40x size increase) because the expensive reserialize happens once per batch, not once per write — and the speedup keeps growing with scale, not just holding steady. Index.add_text() shows the same shape (up to 33.5x at 40,000). .bulk() is opt-in and changes nothing by default — every call remains immediately durable unless you explicitly defer. Reproduce: python bench/scale.py --sizes 1000 5000 10000 20000 40000 --n-writes 50; raw output in bench/pinned/.

Real protocol round-trip: MCP and HTTP, not just the Python API

The numbers above measure the in-process Python API. What a caller actually experiences going through MCP or HTTP includes real subprocess/socket overhead on top — measured with a real python -m rmbr subprocess talked to over real stdio by the real mcp client SDK, and a real uvicorn server on a real OS socket hit with a real httpx client (not the in-process shortcuts the test suite uses for speed), real default embedder, 500-item corpus, 50 samples per call:

Protocol overhead on top of the raw Python API numbers above is small — low single-digit milliseconds, not the dominant cost. session.initialize() (spawning the MCP subprocess and completing the handshake) is the one genuinely slow one-time cost, at ~741ms — pay it once per session, not per call. Reproduce: python bench/mcp_latency.py / python bench/http_latency.py; raw output in bench/pinned/.

Why bge-small-en-v1.5 is still the default

We tested. bench/quality.py measures recall@1 on 150 hand-written (query, correct passage, distractors) examples — 50 each spanning remembered preferences, documentation, and code, the actual shapes of content rmbr indexes — against every same-size-class local embedding model fastembed supports, plus bge-base-en-v1.5 as a "what does 3x the size buy you" reference point:

Nothing in bge-small's own size class beats it with any real confidence — the alternatives above land within about a point of it, which is noise at this sample size. The only model that wins by a real margin is bge-base-en-v1.5: +7.3 points recall@1, at a real, measured cost — 3x the download (210MB) and ~3.9x the per-embed latency (7.5ms vs 1.9ms p50, both still small in absolute terms). We tested that tradeoff and kept the smaller, faster model as the default; if you want the quality bump and don't mind the size, it's a one-line change:

from rmbr.embed import FastEmbedEmbedder
mem = Memory("agents.db", namespace="assistant", embedder=FastEmbedEmbedder(model_name="BAAI/bge-base-en-v1.5"))

Full data and every candidate's per-category breakdown: python bench/quality.py --models candidates.

Roadmap

• v0.1 — Memory + Policy + Index (hybrid BM25 + vector search, metadata filtering), embedding + semantic query caches, MCP support (namespace-pinned), 3-OS CI (Linux/Windows/macOS), true batch ingestion with per-stage timings, async API surface (a-prefixed methods), a Python-aware chunker (stdlib ast, no added dependency), one hosted embedding provider (OpenAI), a 150-example quality eval that confirmed the default embedder against local alternatives, real single-call and bulk benchmark numbers, PyPI trusted publishing, a uvx-launchable console script, and a listing on the official MCP registry
• v0.2 — similarity-based memory dedupe/update (dedupe_threshold), bounded retention (max_memories, forget_older_than), recency-weighted ranking for both Memory.recall() and Index.search(), richer where= filtering ($gt/$gte/$lt/$lte/$in/$nin/$ne, not just equality, now also usable on Memory.list()), a real confidence gate on raw cosine similarity (min_similarity, plus hit.bm25_score/hit.vector_score on every result), an optional local cross-encoder reranker (rerank=True), a conversation-memory convenience (remember_turn()), tool-calling export for hand-rolled agent loops (as_tool()/as_tools(), OpenAI- and Anthropic-shaped, exposing the full where/min_similarity/rerank knob set — not just query/k), LangChain/LlamaIndex retriever adapters (as_langchain_retriever()/as_llamaindex_retriever(), both optional/lazy-imported), two more hosted embedding providers (VoyageEmbedder, CohereEmbedder — same Embedder protocol as OpenAIEmbedder), and two more auto-detected chunkers (split_json, split_rst, both stdlib-only)
• v0.2.1 — adoption/DX polish: a py.typed marker (mypy/pyright now trust rmbr's type hints), README badges (PyPI/CI/license/Python versions), pinned rerank=True/recency_weight latency numbers alongside the existing remember()/search() table, a bench/latency.py fix (each scenario now runs in its own subprocess — running them in one process was polluting each other's tail-latency numbers), and a runnable multi-agent support example (examples/multi_agent_support/). Hardened against real-world tool-calling failure modes surfaced by stress-testing the example against a small, fast, unreliable model: ToolSpec.call() now validates arguments against the tool's own schema and raises a clear ToolCallError instead of a bare TypeError when a model hallucinates one; every built-in tool schema sets additionalProperties: false; to_anthropic()/to_openai() gained a strict=True option; Memory/Index gained stats() and integrity_check() for inspecting a .db file's health without hand-writing SQL; and remember(..., pinned=True) exempts specific memories from max_memories' otherwise-pure-recency eviction
• v0.2.2 — Glama.ai MCP directory listing (verified live, deployed against a pinned commit), an MCP resource template (rmbr://examples/{pattern}, plus rmbr://examples as an index) serving short runnable snippets for common usage patterns to any MCP client that can browse resources, and a fix for serve_mcp() reporting an empty version string in serverInfo (caught live while smoke-testing the Glama deploy)
• v0.2.3 — per-parameter JSON Schema description fields on every MCP tool argument (search/recall/remember's query/k/text/pinned), fixing a real gap Glama.ai's own quality scoring caught: a tool-calling model sees the JSON schema, not the docstring, and none of these parameters had one
• v0.2.4 — two new framework adapters (a real LangGraph BaseStore via as_store(), verified against langgraph-checkpoint's actual op contract; a mem0-API-compatible Memory drop-in reimplemented from scratch, no mem0ai dependency), an optional HTTP server (serve_http/build_app — Starlette+uvicorn, zero new dependencies since mcp already pulls both in, namespace-pinned like MCP, opt-in auth), Memory.get()/Memory.update() for direct record access by id, mem0 added to the bench comparison lane with pinned numbers rerun on the project's bench box, and a real CI/CD hardening pass: a ruff lint gate, genuine subprocess/socket integration tests (a real python -m rmbr MCP subprocess and a real uvicorn socket, not in-process shortcuts), Dependabot, CodeQL, and a SECURITY.md
• v0.2.5 — Memory.bulk()/Index.bulk(), fixing a real O(n)-per-call cost: usearch has no incremental on-disk save, so every remember()/add_text() was re-serializing the entire vector index every time; .bulk() defers that to one write per batch instead (opt-in, default behavior unchanged) — measured on the project's bench box at up to 35.6x faster for sequential writes into a 40,000-item namespace, with the speedup growing as scale grows. PDF/DOCX ingestion for Index.add_files() (rmbr[pdf]/rmbr[docx], optional and lazily imported, loud ImportError instead of a silent skip if the extra's missing). Three new benchmark scripts (bench/scale.py, bench/mcp_latency.py, bench/http_latency.py) measuring real MCP-subprocess and HTTP-socket round-trip latency, not just the in-process API. A second round of Glama.ai MCP quality fixes: real ToolAnnotations (read_only_hint/destructive_hint/idempotent_hint/open_world_hint) on all three tools for the first time, and tool descriptions rewritten to disclose what the JSON schema can't — explicit search-vs-recall usage guidance, k's silent-clamp-not-error behavior on overflow, and remember's max_memories eviction consequence and pinned's permanence. Plus a documentation pass: the version callout and roadmap were stale by two releases, agents.db's auto-creation was never actually stated, and the stale test count was corrected.
• v0.2.6 — fixed Memory(embedder=None)/Index(embedder=None) (the default) constructing a brand-new FastEmbedEmbedder — a fresh fastembed.TextEmbedding/onnxruntime InferenceSession — on every call, with no sharing across instances. Apps opening one Memory/Index per namespace against a shared .db file (the pattern Policy.allow(read=[...]) exists to support) piled up redundant onnxruntime sessions per process, which could reliably crash the process (native heap corruption, worse when another native library shared the process). make_embedder() now shares one FastEmbedEmbedder per model name via a lock-guarded module-level cache, so the default path is safe without callers needing to pass a shared embedder in explicitly. Reported and diagnosed in #18.
• v0.2.7 — fixed a second FastEmbedEmbedder-sharing crash, this time in AnnIndex itself: usearch (>=2.9, confirmed through 2.26.0) leaves a tombstoned node in its HNSW graph after remove(), even once the index is back down to zero vectors — serializing that state and reloading it in a fresh process (exactly what happens the moment a second Memory/Index opens the same .db file/collection after any prior remember()+forget(), add_text()+delete(), or dedupe-triggered update) segfaulted the next add() on that reload, unrelated to the embedder sharing itself despite surfacing in the identical "one Memory per namespace" pattern as #18. AnnIndex now rebuilds itself from its surviving vectors before every serialize whenever a remove() happened since the last one, so a reloaded index never carries a tombstone into a fresh process. Reported and diagnosed in #20.
• Known gaps — none carried over; nothing new opened yet
• Next — a pluggable consolidation hook (mem.consolidate(extractor)): rmbr still never calls an LLM itself, but a caller-supplied extractor callable would let rmbr orchestrate mem0-style fact extraction/dedup/update against your own model choice, without rmbr owning an API key. Deliberately not being built yet. A generic memory-import tool (parsing JSON/YAML/MD exports from other systems) was considered and explicitly deferred — rmbr's remember() is already the universal primitive that job needs, the same way SQLite ships no import tooling for other databases; revisit only for a specific, named source format with real demand, not "agents in general."

License

MIT