open-memex
Persistent memory for your AI coding agents — on your machine, in plain Markdown, shared by every tool you code with.

中文文档
Every AI coding session starts from zero: you re-explain the project, the agent
rediscovers the same gotchas, and yesterday's decisions vanish when the chat
ends. open-memex gives your agents a memory that survives the session. Say
"remember: we deploy on Fridays" once, and next week Copilot, Cursor, opencode,
or Claude Code already knows — because they all read and write the same local
memory on your machine.
- Free and open source (Apache-2.0). No account, no cloud, no telemetry —
everything lives on your machine, in files you can open and edit.
- Local-first: memories are plain Markdown files (the source of truth) with
a rebuildable SQLite keyword index. Nothing leaves your machine unless you
explicitly share it.
- One memory, every agent: wire up several editors with one command; they
share the same memory instead of keeping separate silos.
- Precise synonym search:
ship finds deploy, 单点登录 finds SSO —
a curated multilingual map of equivalents is indexed with every memory at
write time, so no related wording misses the right answer. Deterministic
and auditable, no embedding model required.
- A map, not just search: a derived knowledge graph connects your project
memories — supersedes chains, shared tags, explicit mentions — and injects
a compiled outline at session start, so the agent sees what's in memory
without searching. Structural edges only, no LLM.

New here? This README takes you from install to a working memory in about a
minute. The concept guide explains the mental model in
depth once you're up and running.
Contents
Quick start
You need Node.js ≥ 22.14 (check with node -v). Then:
npm install -g open-memex
open-memex init
init detects the editors you have installed (VS Code, Cursor, opencode, and
Visual Studio when your project has a solution file) and connects each one to
open-memex. Restart your editor afterwards.
See it work (30 seconds):
open-memex add "This project deploys on Fridays"
Now open a new chat in your editor and ask your agent: "When does this project
deploy?" It already knows — no re-explaining. That round trip, capture once and
recall forever, is the whole product. Everything below is detail.
Optional but recommended — check everything is wired up:
Core concepts
Four ideas explain almost everything open-memex does.

1. Two scopes: project and personal.
Every memory belongs to one of two places:
- project — knowledge about one codebase (decisions, constraints, lessons).
Scoped to the current repo automatically; you never set this up by hand.
- personal — knowledge about you (preferences, habits) that applies in
every project. It lives only on this machine and can never be shared into a
repo.
Facts about you ("I prefer concise diffs") go to personal; everything else
defaults to the current project.
2. Capture → recall.
Memories are saved as small Markdown files, one fact each. On the first turn of
every new session, open-memex hands your agent the most relevant ones
automatically, so it starts the session already knowing them. The agent can also
search the full memory on demand. You never have to "load" anything yourself.
3. Your files, your rules.
The Markdown files are the source of truth — open them, edit them, delete them,
grep them. The SQLite index next to them is just a search accelerator and
rebuilds from the files at any time (open-memex reindex). Team sharing, when
you want it, goes through the same review flow as code: nothing is shared
automatically (see Team workflow).
4. Knowledge stays current.
Facts change: decisions get reversed, tools get replaced, yesterday's gotcha
gets fixed. open-memex doesn't let the old version quietly linger —
memory_supersede replaces a memory with its newer version and keeps the
chain, so retrieval serves the current fact while the history stays auditable.
Merely stale memories can be deprecated, and open-memex audit reports
duplicates, stale entries, and broken chains. Pruning is a feature, not
housekeeping: shared memory rots without it.
Which editors, which features
open-memex talks to editors two ways: a native opencode plugin, and a
standard MCP server that any MCP-capable editor can use. (MCP — Model
Context Protocol — is the open standard editors use to give agents extra tools;
open-memex appears in your editor as a set of memory_* tools.) What you get
depends on which path an editor uses:
| Editor | Setup | Tools | Session-start recall | Keyword auto-capture |
|---|
| opencode (native plugin, recommended) | open-memex init --client opencode --global | 6 core tools | Built in — first turn of every session | Yes — remember …, 记住… |
| VS Code (Copilot) | open-memex init --client vscode | All 12 via MCP | Via MCP guidance* | No — the agent saves when you ask |
| Cursor | open-memex init --client cursor | All 12 via MCP | Via MCP guidance* | No — the agent saves when you ask |
| Claude Code | claude mcp add open-memex -- open-memex mcp | All 12 via MCP | Via MCP guidance* | No — the agent saves when you ask |
| Visual Studio 2022 17.14+ / 2026 | open-memex init --client visualstudio | All 12 via MCP | Via MCP guidance* | No — the agent saves when you ask |
| Codex and other MCP clients | open-memex mcp --print-config | All 12 via MCP | Via MCP guidance* | No — the agent saves when you ask |
* MCP has no hard session-start hook, so open-memex sends the agent guidance in
the MCP handshake (including how many drafts are waiting) and init writes the
fuller version into the editor's instruction files. In practice agents follow
it; the opencode plugin is the only path with true built-in first-turn
injection. The Tools section lists the 6 core tools
and the 6 extra workflow tools, so the "6 vs 12" split is explicit.
All editors on the same machine read and write the same memory — a
constraint captured in VS Code is respected in opencode; a lesson learned in
Cursor shows up in Claude Code. (Different machines do not sync automatically;
see the FAQ.)
init also installs an Agent Skill (a short instruction file that teaches
skill-aware agents to use the CLI) for VS Code and Cursor, and for opencode in
per-project MCP mode. With the opencode native plugin wired, no skill is
installed there — the plugin already provides the memory tools, and a second
instruction set only made agents chatty.
How it compares
| open-memex | Instruction files (CLAUDE.md, AGENTS.md, …) | Cloud memory services | Chat history |
|---|
| Where it lives | Your machine + your repos | In the repo | Vendor servers | Gone when the chat ends |
| Who maintains it | Captured as you work; you review | You write and update by hand | The service | — |
| When facts change | Superseded, not silently stale — retrieval serves the current version, chain preserved | You hand-edit; stale copies linger | Varies; often silently stale | — |
| Works across AI tools | Yes — any MCP client (same machine) | One file per tool convention | Per-integration | No |
| Review before sharing | Yes — outbox + pull request | Yes — it's just files | Varies | No |
| Human-readable | Plain Markdown files | Yes | Dashboard / API | No |
Instruction files are great for a handful of standing rules — keep using them
(open-memex can even draft one from your memories; see distill-agents below).
open-memex covers the growing pile of decisions, lessons, and preferences that
no one remembers to write down.
Installation
Requirements
- Node.js ≥ 22.14 (
open-memex doctor verifies this for you). The floor is
the SQLite driver's: better-sqlite3 13 is built against Node-API 10, which
Node gained in 22.14.0 — older Node segfaults on the first database open.
Install the CLI
npm install -g open-memex
That's the stable release. Installing the package may print a reminder to run
open-memex init — the editor wiring is a separate step (see
Setting up your editor), so don't worry if you
don't see the reminder; just run init next.
Two alternatives:
- No install — run via npx:
npx -y open-memex <command> runs any command
without installing (e.g. npx -y open-memex init --client vscode). Slower to
start, nothing to uninstall.
- Alpha builds (newest features, rougher edges, for testers):
npm install -g open-memex@alpha. Check what's published with
npm view open-memex version (stable) and npm view open-memex@alpha version
(alpha).
If open-memex isn't found after installing, your PATH needs attention — see
Troubleshooting.
From source (for contributors)
git clone -b main https://github.com/stoneskin/open-memex.git
cd open-memex
npm install
node --experimental-strip-types src/cli.ts <command>
Setting up your editor
Run init from your project root (the top folder of the repo you're working
in) so the project scope resolves to that repo:
open-memex init --yes
npx -y open-memex init --yes
--yes accepts the recommended defaults for everything init asks about
(editors to wire, auto-capture, first-turn recall). Leave it off if you want to
answer each question. With no --client, init detects your installed editors
and wires them all — one init covers every project. Prefer a single editor?
Pass --client:
VS Code (Copilot):
open-memex init --client vscode
Writes the MCP server entry; reload the window afterwards and confirm the
open-memex server is started in Copilot Chat's MCP panel. By default this
writes a project-level .vscode/mcp.json (an interactive run asks which
level you want); add --global for VS Code's user-level config instead —
one setup that works in every project.
Cursor:
open-memex init --client cursor
Same shape as VS Code: project-level .cursor/mcp.json by default,
user-level MCP config with --global, plus Copilot-style instructions.
opencode (native plugin — recommended):
open-memex init --client opencode --global --yes
Merges the native plugin into your user-level ~/.config/opencode/opencode.json
(or opencode.jsonc if that's the file you already have) — one-time, every
project picks it up, no per-project init. You get the 6 core tools, keyword
auto-capture, and first-turn context injection. (A config file with comments is
left untouched — init prints the line to add by hand.)
Both opencode generations are supported from the same install: init writes
the opencode 1 spelling ("plugin") and the opencode 2 spelling ("plugins")
— each host reads its own key. Opencode 1 needs version 1.18.29 or newer
for this. open-memex doctor tells you if the wiring and the installed host
version don't match.
opencode (as a plain MCP consumer):
open-memex init --client opencode
Writes a project-level opencode.jsonc with the MCP server. Only needed if you
prefer plain MCP over the native plugin — you give up keyword capture and
built-in injection.
Claude Code (from your project root):
claude mcp add open-memex -- open-memex mcp
Visual Studio (from your solution directory):
open-memex init --client visualstudio
Writes solution-level .mcp.json. Requires Visual Studio 2022 17.14+ or Visual
Studio 2026 (Windows-only). Visual Studio also auto-discovers .vscode/mcp.json
and .cursor/mcp.json, so the VS Code setup above works too.
Codex: no init client yet — add the server manually via
open-memex mcp --print-config as a starting point ([mcp_servers] in
config.toml, or codex mcp add).
One-time setup for all projects (VS Code / Cursor)
open-memex init --client vscode --global --yes
Two different "globals" — don't mix them up.
npm install -g open-memex installs the package globally: it puts the
open-memex command on your PATH.
init --global writes the editor config at user level instead of the
project: init once, the wiring works in every project. It works the same
whether the package was installed globally or run via npx.
The --global form writes the server entry to the editor's user-level MCP
config (%APPDATA%\Code\User\mcp.json on Windows,
~/Library/Application Support/Code/User/mcp.json on macOS,
~/.config/Code/User/mcp.json on Linux; ~/.cursor/mcp.json for Cursor)
instead of the project — init once, the server starts in every project.
A per-project .vscode/mcp.json still wins if a project defines its own.
If the user-level file has comments in it (editors accept JSONC), init leaves
the file untouched and prints the exact snippet to paste in by hand.
init behavior notes
- Existing config files are merged, never overwritten — re-running
init
is safe. --force rewrites our entries.
- With no durable
open-memex on PATH (e.g. one-shot npx), init writes an
npx -y open-memex mcp server command into the config so the setup keeps
working. npm i -g open-memex + open-memex init --force switches to the
faster direct command later.
- On an interactive terminal,
init shows the detected editors and asks you to
confirm; scripts and CI never prompt and wire every detected editor.
- If you run bare
open-memex on a machine where init never completed, it
offers to run it for you (interactive terminals only).
Remove the wiring
open-memex uninstall --yes
Reverses init — removes the MCP server entry, the opencode plugin line, the
Agent Skill, and the open-memex section of the editor instructions. With no
--client it cleans up every detected editor; --global limits the cleanup to
user-level wiring. Your memories are never touched.
What init changes on your machine
Everything init writes, in one place:
- Memory data (created on first use, not by
init itself):
%APPDATA%\open-memex\ on Windows, ~/.local/share/open-memex/ on
macOS/Linux — your memory files and the search index. Nothing here is ever
modified by uninstall.
- Editor wiring (removed by
open-memex uninstall):
- opencode: a
"plugin" entry (opencode 1) and a "plugins" entry
(opencode 2) merged into
~/.config/opencode/opencode.json (or .jsonc). Stale my-o-memory
entries from before the rename are removed at the same time.
- VS Code / Cursor: an
open-memex server entry in the user-level or
project-level MCP config, plus an open-memex section in the Copilot
instructions (user-level ~/.copilot/copilot-instructions.md by default;
--instructions project writes .github/copilot-instructions.md in the
repo instead, for teams where everyone uses open-memex).
- Visual Studio:
.mcp.json next to your solution.
- Agent Skill: a
skills/open-memex/ folder for VS Code
(~/.copilot/skills/), Cursor (~/.cursor/skills/), or opencode in
per-project MCP mode (~/.config/opencode/skills/).
- Your repo: nothing. Files only appear in a repo when you explicitly run
submit (see Team workflow) — a local commit, never an automatic push.
Files with comments (JSONC) are never rewritten: init prints the exact
snippet to paste instead.
Capture: how memories get saved
Three ways memories get in:
- Keyword triggers (opencode native plugin only): say
remember …,
note that …, don't forget: …, TIL …, save this … — or in Chinese
记住… / 记一下… / 记录一下… / 别忘了:… — and the sentence is captured
without any tool call. Captures land in the current project by default;
phrases that signal "this is about me" — remember for me …, help me remember: …,
记住我…, 替我记…, 帮我记…, 我觉得…, 我喜欢… — go to personal instead, and
team-context phrases (我们决定…, 帮我们记住…) stay in project.
Two rules keep the noise down (D67): a trigger must be a statement to the
store, so the narration forms 记得… / remind me to… / 别忘了带伞 /
don't forget the wifi password never fire (add a separator — 别忘了:…,
don't forget: … — or that to make it an instruction); and a captured body
under 3 characters is rejected as a fragment rather than saved — capture --dry-run says so instead of dropping it silently. A trigger also owns its
own sentence: the plugin tells the agent what it just stored, and a
memory_add that re-saves the same words within the next few minutes is
refused with the stored id (D73) — one thing you said is one memory, not
three copies of it. A keyword match over the 2000-character hard cap is
dropped, never truncated into a memory — the agent is told in its session
context (D94) and can re-save a shortened version.
- The agent saves it: in any editor, ask your agent to remember something
(or it saves on its own when you state a fact worth keeping) — it calls
memory_add. The routing above is a heuristic; you can always say "save this
to my personal memory" or use the CLI with --scope to be explicit. Long
bodies are never refused: over 500 characters the save proceeds and the
result just carries a soft hint that shorter, self-contained statements
retrieve better (D94).
- Checkpoint proposals: when a task wraps up, the agent proposes 1–3 short
candidate memories distilled from the session — decisions and their reasons,
conventions, gotchas, approaches tried and abandoned — and saves only the
ones you approve. Nothing is written silently: no draft is created behind
your back.
- The CLI:
open-memex add "…" with optional --scope / --tag / --type
/ --aliases.
Aliases. Each saved memory can carry up to 4 alternate phrasings
(synonyms, another language's equivalent) that are indexed with it, so a
question worded differently still finds the memory — "vacation days" finds the
holiday policy. Always on — the agent may attach up to 4 per memory, and you
can pass your own via open-memex add --aliases.
Three ways to add them: open-memex add "text" --aliases "phrase one; phrase two" (semicolon-separated, up to 4); ask the agent ("remember: Fluffy is my
AI agent, alias it as Momo"); or edit the memory's aliases: frontmatter
directly — Markdown is the source of truth and sync picks it up.
Separately, a curated synonym/translation map is expanded into the index
automatically at write time (D80) — ship finds deploy even with no
aliases on the memory. Personal nicknames belong in aliases; shared
dev-domain vocabulary belongs in the curated map. See Retrieval: how
memories come back.
Redaction. Wrap anything sensitive in <private>…</private> and it is
stripped before saving. Recognized secrets (API keys, tokens, high-entropy
credentials) are masked in place — the first 4 characters are kept so you can
tell which key it was, the rest is replaced — and the memory is still saved.
Preview exactly what a message would capture, safely, any time:
open-memex capture --dry-run "…"
If a secret slips through anyway, open-memex forget <id> deletes the memory.
The MCP server exposes thirteen tools; the opencode native plugin exposes the
seven core ones (marked ●). The other six are the team-review workflow tools —
they only matter once you share memories through Git.
| Tool | What it does |
|---|
● memory_add | Save a fact, preference, decision, note |
● memory_search | Keyword search (BM25) across project + personal memories |
● memory_list | List memories in a scope (project, personal, or both) as a numbered inventory; include=all is the audit view |
● memory_supersede | Replace a memory with a newer version (keeps a supersede chain) |
● memory_forget | Delete a memory by id (soft=true hides it instead — retracted, one-way) |
● memory_workingstate | Read/write/clear the per-project working-state note (§3.6, D89) — resumes an interrupted task; scratch state, never a memory |
● memory_useful | Mark a memory as useful (D77) — explicit confirmation only; exposure is counted separately as retrieved_count |
memory_status | Show the sync queue (outbox drafts, repo review states, uncommitted files) plus the D77 reuse report |
memory_submit | Curate deterministic fixes, then move named drafts into the repo memory dir (local branch + commit) |
memory_propose | Copy personal memories into the project scope as review candidates |
memory_promote | Advance proposed → approved → published (or reject / resubmit) |
memory_resolve | List conflicted memory files / 3-way-merge one of them |
memory_pr_status | Map the branch PR's GitHub state onto each memory's review state |
So an MCP-connected editor always has the full set; opencode's plugin covers
capture and recall, and anything workflow-shaped goes through the CLI or an
MCP-connected editor.
Usefulness is measured by reuse, not storage (D77). Every memory carries
local counters: retrieved_count — how often it was actually handed to the
agent (search results, lists, the session-start injection) — and useful
counts, split into useful_self (personal memories, unpublished project
drafts) and useful_team (approved/published project memories). Only
explicit signals raise the useful side: memory_useful when a person
confirms it paid off, a human approval at review, or a supersede. Passive
exposure never counts as useful — a memory that merely showed up in results
has shown nothing yet. Same-side CLI surface: open-memex useful <id...>.
Counters live in the local SQLite index only (never in the Markdown files,
never off this machine). The report rides memory_status (and
open-memex sync-status): per-scope retrieved/useful rates with the self/team
split shown separately, the most-reused memories, and the "zombies" —
retrieved but never confirmed useful, i.e. the decay/prune candidates
(D77-L3). Superseded/deprecated entries stay visible as reuse history but
carry a lifecycle tag, never presented as current guidance.
Memory types
Every memory has a type (what it is) and tags (what it's about). Eleven
types are built in:
| Type | Captures |
|---|
fact | A stable true statement about the project or world |
preference | How someone likes things done |
decision | A choice that was made — the why and the trade-off |
constraint | A rule that must not be violated |
todo | A commitment to do something later |
knowledge | Durable domain or architecture knowledge |
howto | A procedure that worked |
gotcha | A trap to avoid |
lesson | What an incident or mistake taught us |
observation | Something noticed, not yet a conclusion |
reference | A pointer to the authoritative doc (no copying) |
--type accepts any string, but sticking to the built-in set keeps session-start
labels, search, and distill-agents output predictable.
Team workflow: sharing memories through Git
Working solo? You can skip this section — everything above is the whole
product for one person. Nothing below ever happens automatically.
Personal notes stay private. Project knowledge, when you choose to share it,
follows an explicit, reviewable pipeline shaped like code review:
capture → outbox (draft, local) → submit → repo (.ai/open-memex/) → PR review → published → recall
- Capture — save decisions, gotchas, lessons as drafts during normal work.
- Review — drafts wait in a local outbox (on your machine, invisible to
git);
open-memex sync-status — or just saying "sync memory" in chat —
shows what's pending.
- Submit — you name the memories; they move into
<repo>/.ai/open-memex/
with a local commit on your current branch. Before copying, submit prints
a report-only graph lint and auto-applies only deterministic cleanup: exact
duplicates are merged through the supersede chain, empty bodies are
archived, and ragged first-line titles are normalized. Near-duplicates stay
report-only; semantic merge/split needs a human. Every change prints its
undo command, and --no-curate skips the cleanup while keeping the lint
report. open-memex never pushes on its own; it prints the push + PR
commands, and an agent holding your explicit yes can carry them out.
- PR review — memories are plain Markdown; reviewers approve, request
changes, or reject through the normal branch/PR process.
- Recall — published memories are injected at session start and searchable
on demand, for humans and agents alike.
Whoever tends the shared memory follows the curator convention:
what to approve, what to send back, and the hygiene rules that keep shared
memory from rotting.
Retrieval: how memories come back
On the first turn of every session, open-memex injects an [OPEN-MEMEX] block
into the agent's context with the most recent project memories (default: top 8)
and your personal preferences (default: top 5). It looks like this:
[OPEN-MEMEX]
User profile / preferences:
- I prefer concise diffs
Project knowledge (my-repo):
- [decision] We deploy on Fridays; the release train leaves at 10:00
Use the `memory_search` tool to look up more. Use `memory_add` to save new facts.
Do not mention this block to the user unless asked.
It's a snapshot, not the whole memory — the agent can call memory_search any
time for the rest. Both top-N counts are configurable (see Config).
For MCP clients this block is delivered as handshake guidance the agent follows;
the opencode plugin injects it directly on the first turn.
The session-start ritual (wake)
open-memex sessions start from what is already in context, not from questions:
- Read the
[OPEN-MEMEX] block injected at session start (opencode) — or the
equivalent guidance in the MCP handshake instructions. It carries the top
project memories, your preferences, the working-state note from a previous
session (if any), and the compiled memory map.
- If the block is empty or doesn't cover the current task, call
memory_search
once with the task's keywords (try 2–3 phrasings, both scopes) before
concluding anything is missing.
- Don't ask the user to re-explain what memory already holds — repeated
"didn't we decide X?" questions are the failure mode open-memex exists to
prevent. Ask only about what neither the block nor search can answer.
The block is injected once per session, not once per turn — mid-session,
open-memex adds nothing to the context unless a tool is called. The same
ritual is written into the bundled open-memex Agent Skill and the MCP
handshake instructions, so an agent that reads only one of the three still
behaves the same way.
What this costs in context (tokens)
Measured, not asserted (D94, inspired by OptMem's published "426-token
prompt"). On the checked-in fixture in scripts/token-cost.ts — 8 project
memories + 5 personal + the compiled graph outline, default config — with
tokens estimated as characters ÷ 4 (the rough ratio for English prose; a
deliberately tokenizer-free method, no new dependency):
| Surface | Characters | ≈ Tokens | When |
|---|
| Session-start injection block | 2,125 | ~530 | once per session (not per turn) |
| All 13 MCP tool descriptions | 8,814 | ~2,200 | once per session, at the MCP handshake |
| Total | 10,939 | ~2,700 | |
These are approximations and they move with your memory sizes and config —
re-run node --experimental-strip-types scripts/token-cost.ts to measure
your own numbers and to regenerate the fixture's. (The script's output and
this table must agree; update both together, never by hand.)
Retrieval evaluation
We measure recall on a checked-in synthetic fixture instead of asserting it:
60 memories + 47 queries (exact terms, paraphrases, synonyms, Chinese,
cross-language hard cases, synonym-only queries, D81 user-defined alias
queries, and a D79 reuse-signal slice) — reproduce with
node --experimental-strip-types scripts/retrieval-eval.ts.
Current baseline, with index-time synonym/translation expansion (D80) and
query-time user alias expansion (D81), feedback weighting off by default:
recall@1 0.66, recall@5 0.98, MRR 0.80.
Small synthetic corpus, so read it as a regression guard and a starting point,
not a real-world claim.
Optional feedback weighting (D79, first slice, off by default) breaks
near-ties among BM25 hits using the D77 reuse counters: a memory with at least
3 recorded retrievals and a higher reuse_rate moves ahead inside the same
25%-score band. It never edits the BM25 score, never crosses a review tier,
never reorders hits with too little telemetry (new memories keep their exact
BM25 slot), and usefulness still comes only from explicit signals — passive
retrieval exposure never counts. Enable with
open-memex config set feedbackWeighting true; search --explain shows every
promotion.
Three layers, each covering what the others miss: BM25 keyword search
(SQLite FTS5 — exact words, fast, explainable via search --explain);
synonym expansion in two halves — (a) index-time (D80: a curated,
checked-in map of dev-domain equivalents and EN↔ZH translation pairs
expanded into the index at write time, so ship finds deploy and
bug finds defect in a plain single-round query; deterministic, no
model download); (b) query-time user aliases (D81: an alias memory carrying
alias:/target: frontmatter defines your vocabulary once per scope —
e.g. Banana Plan → payment system refactor project — and a query mentioning either side
triggers a second round on the other side, both directions, merged evenly
with round one so alias hits are never down-weighted; no reindex when the
alias changes); and semantic vectors
as an opt-in experiment (local multilingual embeddings measured at R@5 92.7%
on the cross-lingual fixture slice — see docs/retrieval-ablation-study.md
§4.7; no embedding model is ever downloaded without your explicit opt-in).
External check on public data: the same pipeline scores recall@5 97.0% /
MRR 0.909 on LongMemEval-S (470 questions, retrieval stage only — no answer
generation, no judge model; this is not the official LongMemEval score). For
reference, agentmemory's published numbers on the same protocol are 86.2% /
0.715 (BM25-only) and 95.2% / 0.882 (BM25+vector). A 22k-memory pooled stress
run drops recall@5 to 40.6% — the lexical ceiling, measured rather than
asserted. Reproduce with
node --experimental-strip-types scripts/bench-longmemeval.ts --data <path>
(dataset: xiaowu0162/longmemeval-cleaned, MIT).
Security & data
- Local-first: everything lives on your machine (
%APPDATA%\open-memex on
Windows, ~/.local/share/open-memex on macOS/Linux) plus the repos you
choose. Zero cloud calls, zero accounts, zero third-party APIs, zero
telemetry.
- Secrets stay out:
<private>…</private> spans are stripped; detected API
keys/tokens are masked in place before saving. Preview with
open-memex capture --dry-run "…".
- Personal never syncs: the
personal scope is this machine only —
excluded from export by default and can never enter a repo.
- Auditable sharing: team memories move only by explicit
submit, travel
through branch/PR review, and every promote transition is appended to the
memory's review_history (who / when / why).
- You own the files: Markdown is the source of truth — inspect, edit, or
delete anything by hand; the SQLite index rebuilds from the files.
Limitations
Honest edges, so nothing surprises you:
- Keyword search, not semantic. Retrieval is BM25 keyword matching, plus a
curated synonym map expanded into the index at write time (
ship finds
deploy, bug finds defect). Paraphrases outside the curated map
don't match — that's what the opt-in semantic layer is for. (Plain
questions are fine — "how do we…" / "请问…" wording is filtered out before
matching, so asking naturally doesn't dilute the results.) No embedding
model is ever downloaded without your explicit opt-in. (Synonym expansions
live in their own index column, separate from your own aliases — but a
query matching only via an expansion can still rank below direct matches;
that's inherent BM25 length-norm behaviour, not a bug.)
- One machine. Editors on the same machine share memory; there is no
cross-machine sync.
export / import bundles (below) move memory between
machines manually.
- A snapshot, not everything. Session-start recall is a top-N snapshot (8
project + 5 personal by default); older memories are one
memory_search
away, but they are not all in context at once.
- MCP guidance is advisory. Outside opencode, proactive capture and recall
depend on the agent following the handshake instructions — there is no hard
session-start hook in MCP. The tools themselves always work when called.
- Capture routing is a heuristic. Personal-signal phrases (
remember for me …, 我喜欢…) go to personal; everything else defaults to the current
project. When it guesses wrong, say the scope out loud or use --scope in
the CLI.
Upgrading
npm install -g open-memex@latest
Your editor configs point at the installed open-memex command, so upgrades
need no re-wiring. After a major upgrade, run open-memex init --force once to
refresh the installed Agent Skill and instruction files with the latest wording.
If you installed from source or moved the package, --force also re-points the
opencode plugin path.
What changed in each version: see the CHANGELOG.
Storage layout
%APPDATA%\open-memex\ (Windows)
~/.local/share/open-memex/ (macOS/Linux; $XDG_DATA_HOME if set)
├── index.db # SQLite FTS5 index (rebuildable)
└── memories/
├── personal/
│ └── <id>.md
└── project__<name>__<hash12>/
└── <id>.md
Each .md file is one memory: YAML frontmatter (id, scope, type, tags, created_at, schema_version, …) followed by the content. You can edit them by
hand — the index re-syncs from the files, and Markdown is always the source of
truth (open-memex reindex rebuilds the index from scratch).
Once you submit, project memories also live as Markdown files under
<repo>/.ai/open-memex/ (configurable via memoryDir), where they travel with
branches and PRs like any other file.
Config
Settings live in ~/.config/opencode/open-memex.jsonc — the opencode in the
path is historical; this one file is shared by every client. Override the config
path with OPEN_MEMEX_CONFIG and the storage root with OPEN_MEMEX_HOME
(the pre-rename MY_O_MEMORY_CONFIG / MY_O_MEMORY_HOME names are still
honored as fallbacks).
Defaults:
{
"maxProjectMemories": 8,
"maxProfileItems": 5,
"injectOnFirstTurn": true,
"keywordCaptureEnabled": true,
"logLevel": "info",
"graphOutline": true,
"graphOutlineLines": 10,
"feedbackWeighting": false,
"memoryDir": ".ai/open-memex"
}
open-memex config prints the effective config (defaults + file). Change a
setting after install:
open-memex config set keywordCaptureEnabled false
open-memex config set maxProjectMemories 12
open-memex config set sync.autoPull true
open-memex config set graphOutline false
open-memex config set feedbackWeighting true
Settable keys: maxProjectMemories, maxProfileItems, injectOnFirstTurn,
keywordCaptureEnabled, captureAliases, logLevel, memoryDir,
graphOutline, graphOutlineLines, feedbackWeighting, and sync.autoPull
(a dotted key that writes into the nested sync object). Full design:
docs/V2-DESIGN.md.
CLI reference
Setup & health:
open-memex init [--client vscode|cursor|opencode|visualstudio]
[--instructions personal|project] [--global] [--force] [--yes]
open-memex uninstall [--client vscode|cursor|opencode|visualstudio] [--global] [--yes]
open-memex config
open-memex config set <key> <value>
open-memex doctor
open-memex audit
open-memex capture --dry-run "remember that I like concise answers"
open-memex mcp --print-config vscode|cursor|claude|opencode|visualstudio
open-memex --help
open-memex <command> --help
open-memex --version
Memory operations:
open-memex add "This repo uses better-sqlite3" --type fact
open-memex search "auth flow"
open-memex search "auth flow" --explain
open-memex list --scope project
open-memex list --scope both
open-memex list --include all
open-memex supersede <id> "Updated content"
open-memex status <id> deprecated
open-memex forget <id>
open-memex forget <id> --soft
open-memex inventory
open-memex inventory --format json
open-memex inventory --format html
Team review workflow (two homes, one per stage):
Project drafts live in the appdata outbox (git-invisible, branch-independent);
only drafts you approve move into <repo>/.ai/open-memex/, where they follow
branches and PRs. Nothing moves without you naming it. In an AI chat with the
MCP server connected, just say "sync memory" (or "同步记忆") — the agent
runs the status check, summarizes the outbox drafts, and asks which ones to
sync. The server also tells the agent on its own: at session start the
handshake reports how many drafts are waiting, and every memory-changing tool
result carries the current count when it is non-zero.
open-memex sync-status
open-memex submit <id...> [--branch <name>] [--base <branch>] [--no-curate]
open-memex curate undo <id>
open-memex pr-status [--apply]
open-memex pull
open-memex push
open-memex export [--scope project|personal|both] [--type T] [--tag t] [--all] [-o <file>]
open-memex import <bundle.tar.gz> [--dry-run]
open-memex distill-agents [--scope project|personal] [--type t1,t2] [--limit N] [-o <file>]
open-memex propose <id...> --to project [--local-approve]
open-memex promote <id> [--reject] [--resubmit] [--note "..."] [--by NAME]
open-memex resolve [id-or-path]
Whoever tends the shared memory follows the curator convention —
docs/CURATOR.md: what to approve, what to send back, and the hygiene
rules that keep shared memory from rotting.
Maintenance:
open-memex where
open-memex scopes
open-memex reindex
open-memex audit
open-memex wstate
open-memex graph build
open-memex graph lint
open-memex migrate --to-v2 [--dry-run]
The CLI runs under Node 22. From a source checkout it uses the built-in
experimental TypeScript loader (no build step); the published npm package ships
pre-compiled JS (npm run build at publish time). From a source checkout,
prefix every command with node --experimental-strip-types src/cli.ts (or
npm run cli -- <command> for simple cases — npm swallows unknown --flag
args, so prefer direct node).
Session working state
Beyond durable knowledge, the agent can leave one ephemeral working-state
note per project (§3.6, D89): task summary, open files, last commands, where
the work stands, next step. Written by the agent at wrap-up via the
memory_workingstate tool — opencode has no Stop hook today, so "wrap-up"
means when the user signals the session or task is ending (the Claude Code
path arrives with the native plugin work). The note surfaces automatically at
session start so an interrupted task can resume; it is scratch state, never a
memory, and expires after 7 days (configurable) — audit reports expired
notes, humans edit the plain file directly.
open-memex wstate
open-memex wstate --clear
Knowledge graph
A derived, rebuildable graph over the current project's memories (D91, §20
deterministic half) — it answers "what is in memory and how is it connected"
without searching, and it never touches the Markdown source of truth. Structural
edges only, no LLM: supersedes chains, derived-from propose copies,
[[wiki-link]] body mentions (EXTRACTED) and bare-ULID mentions (INFERRED),
and shared-tag co-membership at a fixed
confidence ladder (1 tag → low, 2 → medium, ≥3 → high). Every edge is marked
EXTRACTED (structural fact) or INFERRED (heuristic) with its receipt.
The graph refreshes automatically on every memory write. A compiled outline
(medium+ confidence components, capped and cut-disclosed) is injected once at
session start next to the working-state note; disable it with
config set graphOutline false, tune its length with graphOutlineLines
(default 10, max 30). Personal scope is never graphed.
open-memex graph build
open-memex graph outline
open-memex graph neighbours <id> [--depth 1-6]
open-memex graph path <a> <b>
open-memex graph lint [--json]
lint reports only — nothing is auto-resolved; a human decides. The same report
is the input for D78's pre-submit curation (D93): submit prints it before the
deterministic cleanup. The graph interface is CLI-only; the MCP surface stays at
its thirteen memory tools.
MCP server
The same memory tools over the Model Context Protocol, via a stdio server —
no host-specific plugin needed. Any MCP client can use open-memex.
open-memex mcp
npx -y open-memex mcp
The project scope is resolved from the process working directory, so configure
the server with cwd set to your project root (init handles this for you).
Note: MCP is request/response — it gives the agent tools, not the
opencode plugin's automatic keyword capture or first-turn injection.
Proactive memory use depends on the agent's instructions: the server sends
session-start guidance in the MCP handshake instructions (including the
live outbox draft count at session start, plus the pending count appended to
memory-changing tool results when non-zero), and init writes the fuller
version into the editor's instruction files. Both are advisory — no MCP
consumer offers a hard session-start hook.
Scopes, in detail
- project — scoped to the current repo, keyed off the git origin URL hash
(so clones of the same repo share a scope), or off the cwd path if there is
no remote. Default for new memories.
- personal — global across all your projects, this machine only, never
synced. Use for personal preferences. (v1 called this
user;
migrate --to-v2 renames it.)
See docs/SCOPES.md for the full scope model: key derivation,
migration, visibility, reserved names. The concept guide
walks through the mental model end to end.
Troubleshooting
Every command dies with no output, or the process crashes (exit 139,
0xC0000005, "Segmentation fault").
Your Node is older than 22.14 and the bundled SQLite driver cannot load on
it. better-sqlite3 13 is compiled against Node-API 10, which Node only gained
in 22.14.0; on anything older require() succeeds and then the first database
open segfaults the process with no diagnostic. open-memex now refuses to load
the driver and says so, but anything already crashing was almost certainly this.
Check with node -v, then either upgrade Node (nvm install 22.14 && nvm use 22.14, or any current 22.x/24.x) or pin the driver down with npm install better-sqlite3@^12.11.1. open-memex doctor reports both the Node floor and a
live driver probe. Known upstream: WiseLibs/better-sqlite3#1514.
open-memex is not recognized / command not found.
A global npm install -g puts the open-memex launcher in npm's global bin
folder. If your terminal can't find it, that folder isn't on your PATH:
- Find the folder:
npm config get prefix
- Windows: the launcher (
open-memex.cmd) sits directly in that folder,
e.g. C:\Users\<you>\AppData\Roaming\npm
- macOS / Linux: it's in
<prefix>/bin, e.g. /usr/local/bin or
~/.nvm/versions/node/v22.x.x/bin
- Add it to
PATH:
- Windows: Settings → System → About → Advanced system settings →
Environment Variables → add the folder to the User
Path → restart
the terminal. Verify with where open-memex.
- macOS / Linux: add
export PATH="$(npm prefix -g)/bin:$PATH" to
~/.zshrc (or ~/.bashrc), restart the shell, verify with
command -v open-memex.
- No admin rights / don't want to touch
PATH? Use the npx form —
npx -y open-memex <command> resolves the package itself and needs no
PATH changes.
EBUSY / EPERM on better_sqlite3.node (Windows).
On Windows a loaded DLL is locked: if the open-memex MCP server is running
(VS Code MCP panel, Cursor, etc.), npm install -g open-memex cannot replace
better_sqlite3.node and fails with EBUSY / EPERM. Stop the MCP server
first (or quit the editor), then re-run the install. If it still fails, delete
node_modules/open-memex and any node_modules/.open-memex-* temp folders
under your global npm root and install again.
init says it left a config file untouched.
Your editor config has comments (JSONC) or invalid JSON, and open-memex never
rewrites files it can't parse safely. init printed the exact snippet to add
by hand — paste it in, and you're done. The same applies to the opencode
config: if it has comments, add the "plugin" line manually.
Something's off — run open-memex doctor.
Checks the Node version, config source, scope resolution for the current
directory, and storage writability; verifies VS Code hasn't disabled MCP;
then boots a real MCP server and runs initialize + tools/list against
it — all thirteen tools must show up. It also reports pre-rename my-o-memory
leftovers if any editor config still references the old package name.
FAQ
Do I need git?
No. Capture and recall work in any folder — without a git repo the project
scope simply keys off the folder path. Git is only needed for the team
workflow (submit / PR review), which is optional.
I use several editors. Do they really share one memory?
Yes — on the same machine. Every wired editor reads and writes the same local
memory; see the capability matrix for what
each editor gets. A decision captured in VS Code is respected in opencode.
I work on two computers (office + home). Does memory sync?
Not automatically — memory is per-machine by design, and your personal scope
never leaves the machine it was created on. To move memory manually, use
open-memex export on one machine and open-memex import on the other.
Project memories shared through Git (Team workflow) travel with the repo, so
cloning the repo on the second machine brings the published project memories
along — your personal ones stay behind, on purpose.
Is it really free? Do I need an account?
Free and open source (Apache-2.0). No account, no sign-up, no telemetry, no
cloud calls. If it can't phone home, there's nothing to phone home to: the
only network open-memex ever touches is your own git remote, when you
explicitly push.
I accidentally pasted a secret into a memory. What now?
open-memex search "<part of it>" to find the memory, then
open-memex forget <id> to delete it. To prevent it next time, wrap sensitive
text in <private>…</private> (stripped before saving) — recognized API keys
and tokens are also masked automatically. Preview any message safely with
open-memex capture --dry-run "…".
Do I need to run init for every project?
No. The data layer needs nothing — the project scope is derived automatically
from your cwd's git remote or path, so memories are namespaced per project
with zero setup. The editor wiring is one open-memex init per machine
(user-level wherever the editor supports it). Run it again only after
upgrading (--force) or if you switch editors.
Does opencode need init?
Two paths. Recommended: open-memex init --client opencode --global — it
merges the native open-memex plugin into ~/.config/opencode/opencode.json
for you. One-time setup, applies to all projects, and additionally enables
keyword auto-capture and first-turn memory injection. Prefer to do it by hand?
Add "plugin": ["file:///absolute/path/to/open-memex/src/index.ts"] (the
installed package's path) to that file instead. As a plain MCP consumer:
open-memex init --client opencode writes a project-level opencode.jsonc
(no hooks). If your user-level config has comments, init leaves it
untouched and prints the manual step.
VS Code — run init once, or per project?
Once. Plain open-memex init auto-detects VS Code and writes the MCP server
entry to VS Code's user-level mcp.json (%APPDATA%/Code/User/mcp.json on
Windows, ~/Library/Application Support/Code/User/mcp.json on macOS,
~/.config/Code/User/mcp.json on Linux), so the server starts in every
project. A per-project .vscode/mcp.json still wins when present, and the
entry keeps cwd=${workspaceFolder} so project-scope resolution keeps
working per window. If your user-level mcp.json has comments (VS Code
accepts JSONC), init leaves it alone and prints the exact snippet to add by
hand. An empty file is treated as blank and written to directly.
How do I remove the editor wiring?
open-memex uninstall reverses init: it removes the MCP server entry, the
opencode plugin line, the Agent Skill directory, and the open-memex section of
the Copilot instructions. With no --client it cleans up every detected
editor; --global limits the cleanup to user-level wiring. Your memories are
never touched.
I upgraded Node, or switched versions with nvm. Do I need to reinstall?
The package itself doesn't care: its SQLite driver is a Node-API prebuild, so
it loads on any supported Node (≥ 22.14) with no recompiling, and your
memories live outside the install. But version managers (nvm and friends)
keep a separate global package folder per Node version, so after a switch
open-memex may simply be "not found". Run npm install -g open-memex once
under the new Node, then open-memex doctor to confirm the driver loads.
Project status
open-memex is stable and in daily use; the current stable line is published on
npm as latest, with alpha builds for testers. Release history lives in
GitHub Releases; design
decisions are recorded in the append-only log at
docs/V2-DESIGN.md.
On the horizon (no version promises): native agent plugins for more editors as
enhancements over the same MCP tools; local embeddings as an opt-in experiment
(no model is ever downloaded without asking); an org layer only if real
multi-repo sharing, ACL, or compliance needs demand it.
License
Apache-2.0