PLUR — Your agents share the same memory

Persistent, open memory for AI agents — local-first, zero-cost, shared across MCP tools (Claude Code, Codex, Cursor, Hermes, OpenClaw). Your agent's memory is plain-text engrams you can read, correct, and delete — not weights you can't.

plur.ai · Benchmark · Engram Spec · npm · Comparisons

Benchmarks

PLUR is memory, not just retrieval — so we measure it on more than one axis, on the full corpus, and we publish the harness so you can reproduce every number.

Retrieval recall — full LongMemEval-S (N=500), R@5, fully local:

Numbers come from plur-ai/plur-bench, which is the source of truth for every benchmark figure PLUR publishes. Where an in-repo number and a plur-bench number disagree, plur-bench wins — it is the reproducible harness, and it is what CI regression-checks.

Chunk granularity, canonical-doc scoring, corpus SHA256 pinned — reproduce it in plur-ai/plur-bench. No cloud call is required for any of these numbers (an optional cloud embedder, openai-3-large, reaches 97.0% hybrid). A faster reranker — ms-marco-minilm-l6 (p50≈245ms vs BGE's ≈5s on CPU) — trades a little recall for sub-second latency.

Run it yourself — and tell us what you get. The harness is plur-ai/plur-bench: CPU-runnable, no API key needed for the local path, corpus auto-fetched and SHA-verified. If you run it, we'd genuinely love to see your numbers — open an issue or discussion with your results, especially if they don't match ours. Independent reproduction is worth more than any number we publish, and we'll gladly credit you.

Retrieval ≠ answer accuracy — and we report them separately, never conflated. End-to-end (LLM-judge) answer accuracy with the reranker stack is 60.5%, versus 52.0% for dumping full context into the prompt and 5.5% with no memory at all.

Agent-task impact — same task, with memory vs without: Haiku + PLUR outperforms Opus without it at roughly 10× less cost; house rules 12–0 across Haiku, Sonnet, and Opus.

Operational — local-first, zero-cost search, data-sovereign by design.

More in progress: LoCoMo, agentic task suites, cross-tool portability, decay / contradiction correctness. Full methodology →

The idea

You correct your agent's coding style on Monday. On Tuesday, it makes the same mistake. You explain your architecture in Cursor. That night, Claude Code has no idea.

PLUR fixes this. Install it once, and corrections, preferences, and conventions persist — across sessions, tools, and machines. Your memory is stored as plain YAML on your disk. No cloud, no API calls, no black box.

The interesting part: in our tool-routing and local-knowledge benchmark, Haiku with PLUR memory outperformed Opus without it — 2.6x better on tool routing, at roughly 10x less cost. Turns out the bottleneck isn't model intelligence. It's context.

The model is rented; your memory is owned. Swap Haiku for Opus for whatever ships next month — the reasoning is a commodity you don't control. The part that's yours — everything the agent has learned about your work, your corrections, your conventions — shouldn't live in someone else's cloud or be baked into weights you can't read. PLUR keeps it in plain files on your disk, in an open format you can inspect, correct, and delete. That's what owning your intelligence actually means.

Install

Tell your agent

Paste this to your coding agent (Claude Code, Cursor, Windsurf, OpenClaw):

Set up PLUR memory for me: run `npx @plur-ai/mcp init`, then check my PLUR status to confirm it works.

Prefer a guided setup? plur.ai has the exact config for your tool — Claude Code, Cursor, Windsurf, or OpenClaw.

Manual setup (Claude Code)

One command sets up everything — storage, MCP config, and Claude Code hooks:

npx @plur-ai/mcp init

This creates ~/.plur/ for storage, adds PLUR to your .mcp.json, and installs Claude Code hooks for automatic engram injection. The hooks also auto-close the memory lifecycle: a SessionEnd hook captures a closing episode and cleans up session state when a conversation ends, so memory closes cleanly even if the agent forgets to call plur_session_end. PLUR is installed globally — one MCP server, one store, available in every project. You only run init once.

For multi-project setups, use domain/scope to separate knowledge:

cd ~/projects/my-app
npx @plur-ai/cli init --domain myapp --scope project:my-app

This creates a .plur.yaml in the project with defaults that hooks apply automatically. Engrams learned in that project are tagged; recall filters by scope but always includes global knowledge.

Set scope per engram, by content. Scope is not a once-per-session setting — every plur_learn call takes its own scope, chosen from what the engram is about. Team/shared knowledge goes to a team scope (e.g. group:<org>/<team>, used by PLUR Enterprise); project details to project:<name>; personal preferences stay local. Don't let team-relevant knowledge fall back to global by omitting scope — global leaks into every project and (with a team store configured) never reaches the team. plur_session_start lists the remote scopes a token can write to.

Global install (faster startup)

npm install -g @plur-ai/mcp
plur-mcp init

Cursor

Run init from your project root — it sets up Cursor's .cursor/mcp.json (plus Cursor hooks and a context rule):

npx @plur-ai/mcp init

PLUR runs under a lean tool profile in Cursor (PLUR_TOOL_PROFILE=cursor) — Cursor caps the tools a workspace can expose, so PLUR surfaces a curated core set (learn / recall / inject / status) instead of all 43, with the rest reachable through plur_admin. Cursor support shipped in v0.13.

Codex

npx @plur-ai/cli init --codex

Registers the MCP server via codex mcp add, writes lifecycle hooks to ~/.codex/hooks.json, and adds a PLUR section to AGENTS.md. Auto-detected when ~/.codex/ exists.

Injection uses hybrid search (BM25 + embeddings) with an automatic BM25 fallback if the embedder is slow or unavailable. Set PLUR_HOOK_HYBRID=0 to force BM25 (applies to the Antigravity hooks too; PLUR_CODEX_HYBRID is honoured as an alias). PLUR_HOOK_HYBRID_DEADLINE_MS tunes the fallback deadline — keep it below your harness's hook timeout (Codex 25s, Antigravity 20s).

One manual step after install: open Codex, run /hooks, and trust the PLUR entries. Codex fingerprints every hook and refuses to run untrusted ones — silently, with no warning and a zero exit code. Until you trust them, memory simply never loads. plur doctor says so too.

Which integration you get

Every MCP client can call PLUR's tools. Only some have an adapter — the hooks and always-on context that make memory load automatically instead of waiting for the agent to think of it. Without one, recall and learning depend entirely on the model choosing to call the tools, which degrades badly under context pressure.

If your harness is in the last row, paste the PLUR section from CLAUDE.md into its own context file (AGENTS.md, GEMINI.md, …) as an interim measure — that restores the instruction layer, though not automatic injection.

Antigravity CLI (agy)

npx @plur-ai/cli init --antigravity

Writes hooks and the MCP server into agy's global config (~/.gemini/config/) and adds a PLUR section to AGENTS.md. Auto-detected when ~/.gemini/antigravity-cli/ exists. No trust step — agy runs configured hooks on first invocation; just restart agy.

Antigravity has no session-start event and no per-prompt hook, so PLUR drives everything from PreInvocation: per-prompt recall is read from the conversation transcript, and the turn's memory is re-injected as an ephemeral message on every model invocation so it survives tool calls without accumulating in history.

Gemini CLI users: Google is transitioning Gemini CLI to Antigravity — install agy and run the command above. Gemini CLI itself remains tools-only.

OpenClaw

openclaw plugins install @plur-ai/claw
openclaw config set plur.enabled true

That's it. PLUR works in the background from here. No workflow changes needed — just use your tools as usual. Corrections accumulate automatically.

DeepSeek Harness

dsh plugin add @plur-ai/dsh

Native, not an MCP bridge. PLUR mounts as a Cordis plugin and writes your engrams straight into the system prompt, so the model reads them the way it reads its own instructions — no tool call, no round trip, and no turn spent deciding whether to look. The section is re-rendered on each assembly rather than appended, so memory does not accumulate in the context as a session runs.

Five tools (plur_recall, plur_learn, plur_forget, plur_feedback, plur_status) are still registered for when the agent wants to reach for memory deliberately. Scope defaults closed — each workspace gets its own, resolved from its .plur.yaml.

/plur reports status; /plur-memory opens the memory viewer below.

Hermes Agent

pip install plur-hermes
npm install -g @plur-ai/cli

The plugin registers automatically via Hermes' plugin system. It injects relevant memories before each LLM call, extracts learnings from agent responses, and exposes all PLUR tools to the agent. Hermes shells out to the PLUR CLI.

Python SDK (LangChain, llama.cpp, scripts)

For Python environments that aren't Hermes:

pip install "plur-ai @ git+https://github.com/plur-ai/plur.git#subdirectory=packages/python"
npm install -g @plur-ai/cli   # bridge (required)

Note: plur-ai is not yet on PyPI — use the git install above until #915 is resolved.

from plur_ai import Plur

plur = Plur()
plur.learn("always use async generators for streaming LLM output")
results = plur.recall("streaming patterns")
context = plur.inject("write a streaming endpoint", limit=10)

plur-ai bridges to the same on-disk store as Claude Code and OpenClaw — memory written from Python is immediately visible across all your tools. See packages/python/examples/ for LangChain and llama.cpp integration examples.

Verify it works

Ask your agent: "What's my PLUR status?" — it should call plur_status and return your engram count and storage path.

Read your memory

plur dashboard

Opens a local page listing every engram: what was learned, what actually gets recalled, and how often. Read-only, loopback-only, and served from your own machine — nothing is uploaded. --port moves it, --no-open skips the browser. Inside DeepSeek Harness the same page is one /plur-memory away.

Available in English and 中文; it follows your browser, or ?lang=zh.

See it in action

Once it's running, teach your agent something once:

"Always use pnpm in this project — npm install breaks the lockfile in CI."

Start a new session the next day and ask:

You: How do I run the tests?

<plur-memory> 1 engram · project:my-api </plur-memory>

Agent: Use pnpm — you mentioned npm breaks the lockfile in CI:

  pnpm test                           # full suite
  pnpm test -- src/auth.test.ts       # single file

New session. No reminder. The correction was there.

That's the moment PLUR pays off — the agent remembers a project convention you mentioned once, without it being in any file it can read.

How it works

PLUR has two storage primitives:

Engrams — learned knowledge that persists across sessions. Each engram is a typed assertion ("always use blue-green deploys", "never force-push to main") with:

• Activation — retrieval strength that decays over time (ACT-R model) and strengthens on access. Stale facts naturally fade from injection without manual cleanup.
• Feedback signals — positive/negative ratings that train injection quality over time
• Scope — hierarchical namespace (global, project:myapp, cluster:prod, service:api) controlling where the engram applies
• Polarity — automatic classification of "do" vs "don't" rules, so constraints are injected separately from directives
• Associations — links to other engrams, including co-access edges that form automatically when engrams are recalled together

Episodes — timestamped event records for "what happened when." Each episode captures a summary, timestamp, agent attribution, and channel. Use episodes for incident timelines, session logs, and operational history. Query by time range, agent, or channel.

You correct your agent  →  engram created  →  YAML on your disk
Agent fixes an incident →  episode captured →  timeline searchable
Next session starts     →  relevant engrams injected  →  agent remembers
You rate the result     →  engram strengthens or decays  →  quality improves
Unused engrams          →  activation decays  →  naturally fade from injection

Search is fully local: BM25 (with IDF weighting, TF saturation, length normalization) + BGE embeddings + Reciprocal Rank Fusion. Zero API calls, zero per-query cost. Benchmark methodology →

Plugins (OpenClaw, Hermes) automatically capture learnings from agent conversations — no manual saving needed. The agent's corrections become engrams without you doing anything.

See the full engram spec for schema details, activation model, and injection algorithm.

Open format

The engram is an open, versioned format — not a black box. Every engram is plain YAML validated against a published JSON Schema, generated from the same Zod source the engine uses (the schemas live in spec/). Read it, diff it in git, write your own tooling against it, or build a different engine on the same format — your memory isn't locked to PLUR.

Usage

import { Plur } from '@plur-ai/core'

const plur = new Plur()

// Learn from a correction. The engine's read and write methods are async —
// they return promises so a `Plur` can be backed by a network store as well
// as by the default local YAML one.
await plur.learn('toEqual() in Vitest is strict — use toMatchObject() for partial matching', {
  type: 'behavioral',
  scope: 'project:my-app',
  domain: 'dev/testing'
})

// Recall (hybrid: BM25 + embeddings, zero cost)
const results = await plur.recallHybrid('vitest assertion matching')

// Inject relevant engrams into agent context. You get context blocks ready to
// paste into a prompt plus the IDs that went into them — not the engrams
// themselves. `budget` is the ceiling in tokens; selection fills it by relevance.
const injection = await plur.inject('Write tests for the user service', {
  scope: 'project:my-app',
  budget: 2000
})
console.log(injection.directives)   // also .constraints, .consider
console.log(`${injection.count} engrams, ${injection.tokens_used} tokens`)

// Feedback trains the system — rate anything you have an ID for, whether it came
// back from recall or went out in an injection (injection.injected_ids).
if (results[0]) await plur.feedback(results[0].id, 'positive')

// Capture an event (episode). Episode operations stay synchronous — they are
// backed by episodes.yaml, not the engram primary store.
plur.capture('Fixed CrashLoopBackOff on bee-3-4 by increasing memory limits', {
  agent: 'claude-code',
  channel: 'terminal'
})

// Query timeline
const incidents = plur.timeline({ agent: 'claude-code' })

// Sync across machines (use a private git remote — all engrams including private-visibility ones are pushed)
await plur.sync('git@github.com:you/plur-memory.git')

Tools

The outbox

A write to a team scope goes to that team's remote store. When the store cannot be reached — VPN off, server down, token expired — the engram is not lost and not silently dropped: it is written locally with queue metadata and retried on the next session start, on plur sync, or on demand.

The queue is not a directory. It lives as structured_data._outbox inside the affected engrams in engrams.yaml, which is why it needs a command to see:

plur outbox            # what is queued, for which scope, how long, last error
plur outbox --flush    # retry now

The same thing is available to agents as plur_outbox ({flush: true} to retry), and plur status reports the pending count. Neither surface prints the target URL or the token.

The memory receipt

plur receipt (and the plur_receipt MCP tool) show what your memory actually did — counted from PLUR's own retrieval history, never estimated:

Your Memory Receipt
===================
  2026-07-03 .. 2026-07-22  (71 sessions)

  423 times a memory you taught PLUR
  was put in front of the model.
  (plus 45 times an installed-pack memory)

  across 71 retrievals in 71 sessions
  drawing on 162 distinct engrams

  MOST-RELIED-ON
      34x  PLUR positioning thesis across every vertical: PLUR layers …
      28x  Datacore app CoS architecture: reasoning layer added on to…
      ...

  STORE HEALTH
       4,517   engrams stored (you: 3,746, packs: 771)
         162   retrieved at least once (4% of store)
       4,355   not retrieved since 2026-07-03 (96%)
    Over a short logging window a low rate is expected, not a fault —
    memory is meant to be selective, and much of the store predates logging.

(REUSE stats and coverage caveats are also shown; trimmed here for length.)

It is local and read-only, and carries no dollar or token figure by design: on a subscription your marginal token cost is zero, and the value of an avoided rediscovery is not measurable from this data. The receipt reports only what it can count. Activation rate is store coverage over the logging window, not a quality score — it is naturally low and falls as you add engrams. --days N narrows the window; --json emits the raw shape. (The plur_receipt MCP tool returns the same figures plus a one-line summary that carries this framing to the agent.)

Syncing across devices

plur.sync(remote) is git underneath: it commits your engram store and pushes it to the remote you give it. What gets pushed depends on the remote's declared type (sync.remote_type in config.yaml, or the remote_type argument):

• personal (default) — your own backup/mirror across your machines. The remote receives everything that is pushed, including visibility: private engrams: private visibility means "don't share this in a pack", not "don't mirror it to my own devices", so private engrams intentionally follow you from machine to machine. Because of that, always use a private git remote (a private GitHub/GitLab repo, or your own server). PLUR surfaces a warning in the sync result whenever private engrams are present. Never point a personal sync at a public repository.
• shared — a team-visible remote. Only engrams with a shared-family scope (group:/project:/space:/team:/org:/public) and a non-private visibility are pushed; personal-family engrams (local, global, user:*, agent:*) and private-visibility engrams never reach the remote, by construction. Note the default visibility is private, so a shared remote receives only engrams whose visibility was set deliberately — teammates get what you chose to share, nothing else. The same guarantee covers the sibling store files (#686): an episode, candidate, or tension record is pushed only when every engram it references is itself in the push set — records derived from personal or private engrams (a tension's statement snapshots, a failure-report episode) stay local, as does any record referencing an engram the filter cannot resolve.

In both modes scope: local engrams are machine-specific by design (paths, local ports, per-host quirks), so they are stripped from every commit and never reach any remote. Stripping happens on the staged blob: your local working copy always keeps every engram.

Benchmark details

Per-category retrieval recall, from an earlier in-repo run — full LongMemEval-S (N=500), fully local (BGE-small + BGE-reranker-v2-m3, chunk granularity). Its overall figure (98.0%) predates the current plur-bench measurement of the same stack (97.6%) and has not been re-run per category; treat the shape as indicative and the headline table above as current.

Retrieval recall (finding the right memory) and end-to-end answer accuracy (whether the model then answers correctly) are different axes — PLUR measures and reports them separately, never conflated. The agent-impact figures above come from a same-task A/B run (memory vs none).

Full methodology →

PLUR vs other agent-memory tools

Mem0, Letta (MemGPT), and Zep solve real problems — a drop-in memory API (Mem0), a self-managing agent OS (Letta), a temporal knowledge graph (Zep). PLUR's bet is a combination none of them ship together:

• Plain-text you own — engrams are human-readable YAML you can read, git diff, edit, and provably delete. Not opaque vectors, agent-state blocks, or graph nodes you need tooling to inspect.
• Local-first, zero-cost — hybrid BM25 + local embeddings, fully offline, no API bill (98% R@5 on the full LongMemEval-S corpus with no cloud call — see above).
• Team-shareable via git — plur sync is git underneath, so the same memory follows you across machines and across a team. Most tools are single-user-local or cloud-team; PLUR is both, and you keep the data.
• Cross-tool — the same ~/.plur/ store works in Claude Code, Cursor, Windsurf, OpenClaw, and Hermes. Your memory isn't trapped in one vendor.
• It learns and forgets — feedback-trained retrieval with ACT-R decay and an on-demand contradiction scan, not a grow-forever store.

If you need a hosted memory API or a temporal knowledge graph, use the tool built for that. If you want memory you can read, own, share with your team, and move between tools, that's PLUR. Side-by-side detail: comparisons/.

What PLUR is — and isn't

PLUR is agent memory — it stores corrections, preferences, conventions, and architectural decisions that an AI agent learns during work sessions, and injects them back when they're relevant.

PLUR is not a general-purpose search engine, a codebase indexer, or a replacement for code intelligence tools. It doesn't parse ASTs, navigate class hierarchies, or search your source files. If you need code-aware search (tree-sitter, language server features, symbol lookup), tools like claude-mem or your IDE's built-in search are the right choice.

The two are complementary:

While search is a core part of PLUR (finding the right engram to inject), the search targets are always engrams — not files, not code, not documents. PLUR's hybrid search (BM25 + embeddings + RRF) is optimized for short natural-language assertions, not source code.

Packages

packages/ui is internal — the memory viewer's pages, bundled into the CLI and the DeepSeek Harness plugin rather than published. It is not on npm.

Architecture

@plur-ai/core
├── engrams.ts           Engram CRUD + YAML persistence
├── episodes.ts          Episode capture + timeline queries
├── fts.ts               BM25 with IDF, TF saturation (k1/b), length normalization
├── embeddings.ts        BGE-small-en-v1.5, 384-dim, local ONNX
├── hybrid-search.ts     Reciprocal Rank Fusion
├── inject.ts            Context-aware selection + spreading activation
├── decay.ts             ACT-R activation decay
├── secrets.ts           Secret detection (API keys, passwords, tokens)
├── sync.ts              Git-based sync + file locking (O_EXCL)
├── storage.ts           Path detection + YAML I/O
└── storage-indexed.ts   Optional SQLite read index

@plur-ai/mcp          Wraps core as MCP tools
@plur-ai/claw          OpenClaw ContextEngine hooks (assemble/compact/afterTurn)
plur-hermes            Python plugin for Hermes Agent (auto inject/learn)
plur-ai                Python SDK — direct learn/recall/inject for scripts and frameworks

Storage

Everything is plain YAML. Open it, read it, edit it.

~/.plur/
├── engrams.yaml     # learned knowledge (source of truth)
├── episodes.yaml    # session timeline
├── config.yaml      # settings
└── engrams.db       # optional SQLite read index (auto-generated)

PLUR_PATH overrides the default location.

Indexing is on by default (index: true) and the backend is chosen from the size of your store, so there is normally nothing to configure:

YAML stays the source of truth in every tier except postgres (ADR-0001, ADR-0005) — the index is a cache that rebuilds automatically, and you can delete it anytime. Set backend: in config.yaml to pin a tier explicitly.

One caveat on the postgres tier, stated here because it is this table's headline row: core does not write embeddings to a Postgres primary store (ADR-0005 amendment). Keyword/BM25 recall runs in SQL, but engram_embeddings stays empty unless your deployment populates it, so semantic and hybrid recall fall back to loading engrams and scoring in memory — correct results, at the O(N) cost this tier otherwise avoids. The adapter says so once at schema init; vectorIndex: 'exact' acknowledges and silences it.

sqlite (engrams.db, via better-sqlite3) is the legacy index and is no longer selected automatically.

Requirements

• Node.js 18+
• 2GB RAM minimum — the embedding model (ONNX runtime) needs ~1GB for installation. On servers with less RAM, embeddings are skipped and search falls back to BM25 keyword matching.

Development

git clone https://github.com/plur-ai/plur.git
cd plur
pnpm install && pnpm build && pnpm test

~3500 tests across ~200 files. pnpm test:watch for development.

Contributing

• Bug reports — issue with reproduction steps
• Feature requests — issue describing the use case
• Code — fork, branch, PR. Tests required.
• Integrations — build PLUR support for other tools

Before submitting: pnpm test passes, pnpm build succeeds, no new external deps in core without discussion.

Conventions: TypeScript, Zod validation, Vitest, no external APIs in core, YAML storage, zero-cost search by default.

License

Apache-2.0