atlas-recall

Deterministic markdown memory and knowledge graph for AI coding agents: index a directory of notes, retrieve the relevant ones, and search/trace the same corpus as a wikilink graph. No vector database required, and no agent lock-in -- the engine is a plain recall CLI, so it works with any agent or workflow that can run a shell command.

Retrieval and search are two interfaces on one engine, not two separate tools bolted together. The same corpus and the same config back both: recall query answers "what's relevant" automatically (wire it into anything that can invoke a command per turn), and recall find / recall map / recall trace answer the same question interactively for a human, plus a wikilink graph over your notes.

The closest neighbour is basic-memory, which also stores knowledge as plain markdown with wikilinks, also builds a traversable graph, also ships an MCP server, and also ships editor hooks. If you want a full local-first knowledge management system, look at it first.

Two differences, both checkable rather than asserted. Weight: this package's default path is one runtime dependency (rank-bm25); basic-memory 0.23.2 declares 43 ungated runtime dependencies, among them fastapi, sqlalchemy, fastembed and sqlite-vec. That is from its own PyPI metadata, so re-check it rather than trusting this line. Trigger frequency: the automatic path here fires on UserPromptSubmit, so retrieval is injected on every turn; basic-memory's hook briefs once at SessionStart, and mid-session recall goes back through a model-invoked MCP tool. Read as "different cadence", not "one of these lacks hooks" -- both ship them.

Two ways to wire it into an agent, both driven by recall install (no copy-pasting JSON by hand): an MCP server (atlas-recall-mcp) that exposes retrieval and the knowledge graph as structured tools over local stdio, and a Claude Code hook (recall hook) that injects retrieval results straight into the prompt on every turn. Any other agent that can run a shell command gets the same engine by calling the recall CLI directly.

Supported clients

Client	How	Config written
Claude Code	UserPromptSubmit hook	~/.claude/settings.json
Claude Desktop	MCP (stdio)	~/Library/Application Support/Claude/claude_desktop_config.json (macOS)
Cursor	MCP (stdio)	~/.cursor/mcp.json
Windsurf	MCP (stdio)	~/.codeium/windsurf/mcp_config.json
Cline (VS Code)	MCP (stdio)	.../globalStorage/saoudrizwan.claude-dev/settings/cline_mcp_settings.json
Codex CLI	MCP (stdio), via codex mcp add	~/.codex/config.toml

The Cline path assumes standard VS Code; recall install --client cline refuses (rather than silently creating a file Cline will never read) if that globalStorage directory doesn't exist -- VS Code Insiders and VSCodium use a different one, so wire it in by hand in that case.

Every path above was verified against that client's own docs, not guessed -- recall install --client <name> writes to it directly (merges, backs up, idempotent, --dry-run to preview). Zed also takes a local stdio MCP server the same way the others do; it just isn't wired into recall install yet, so point it at atlas-recall-mcp by hand.

Not supported: ChatGPT. ChatGPT's web and desktop connectors need a remote HTTP MCP endpoint, not a local stdio process -- exposing a user's private notes over a network endpoint is a deployment decision for whoever runs this, not something this package does on its own, so it isn't shipped here.

The problem

AI coding agents forget everything between sessions unless you paste context in by hand. The usual fix is a vector database: embed every note, embed the query, take the top-K by cosine distance. In practice that ceiling doesn't discriminate well on a personal notes corpus (tens to low hundreds of files), it needs an embedding model running somewhere, and it's one more service that has to be up before the first command works.

The measured result

A dense-only retriever at a swept 0.45 cosine cutoff, measured against a 15-query labeled should-hit set over a 108-document corpus, admitted an average of 16.1 of 108 documents per query at 100% hit-recall (14.9% of the corpus) - the cutoff wasn't discriminating anything, it was a junk floor. Swapping the embedding model and re-sweeping the cutoff (0.48, still dense-only) brought that down to 2.9 of 119 documents at the same 100% hit-recall (2.4% of the corpus), roughly 5.5x fewer false admits.

The retriever this package ships (retrieval.py) goes further: BM25 lexical search fused with the dense path via Reciprocal Rank Fusion (RRF), gated by a conservative admission floor, scored by recency decay and a light priority tiebreak. RRF fuses by RANK, not by raw score, which is what lets a keyword-shaped query win on BM25 and a paraphrase win on dense without needing a BM25 score and a cosine distance to live on the same scale. The 5.5x number above is the dense half of that measurement, reproduced honestly here because it's the reason RRF replaced a plain distance cutoff rather than the whole story - re-run recall query against your own corpus and judge the fused result on your own notes; a number measured on someone else's corpus is a reason to expect a good default, not a guarantee.

Measuring retrieval quality on your own notes

The 16.1-vs-2.9 number above is a claim, not something this package gives you a way to check, until now. recall eval runs a labeled query set against your own configured corpus and reports precision@k and hit@k for two arms: dense-only (raw top-k by distance, no cutoff, no fusion -- the naive baseline the 16.1/2.9 numbers above measured), and this package's actual default (BM25, fused with dense via RRF when it's configured). With no arguments it runs against a small built-in synthetic fixture corpus (not your notes) so the command produces real output immediately after install:

$ recall eval
recall eval -- corpus: built-in fixture corpus  (k=3, n=22 labeled queries)

query                                                      expected                                 dense@k  hybrid@k
----------------------------------------------------------------------------------------------------------------------
espresso grind size fine or coarse                         reference-espresso-grind-size            miss     HIT
...
favorite jazz trumpet solo recordings                       (none)                                  HIT      HIT

dense-only  (raw top-k by distance, no cutoff, no fusion)  precision@3=13.64%  hit@3=13.64%
hybrid RRF  (this package's default)  precision@3=73.48%  hit@3=100.00%

dense-only arm returned nothing for every query -- dense retrieval isn't
configured/reachable for this corpus (needs `recall index --dense`, the
`[dense]` extra, and a running Ollama). The hybrid numbers above are this
package's real BM25-only behavior on this corpus.

That 13.64% for dense-only is NOT "dense retrieval is bad" -- it's three negative controls correctly scoring as correct rejections (an empty result is the right answer when nothing should match) with dense disabled in this demo run, not a real dense-only measurement. Enable recall index --dense on your own corpus to get a real dense-only number to compare against.

Point it at your own corpus with --labels your-labels.json (a JSON array of {"query": ..., "expected": "<note-name>-or-null"} pairs, null for a negative control your corpus shouldn't match anything for):

$ recall eval --labels my-labels.json

What this does and doesn't prove, stated plainly:

• You CAN measure, on your own notes: your own BM25-only hit rate and precision, the delta from turning recall index --dense on, whether min_score is admitting junk at your corpus's size, and whether dense_dist_max needs recalibrating for your embedding model.
• You CANNOT reproduce the 16.1-vs-2.9 headline above from this harness. That number came from a separate offline cutoff sweep across two embedding models (nomic-embed-text, then bge-m3), not from anything recall eval runs. Don't read a good recall eval result as confirming that specific number; read it as evidence about YOUR corpus.

The dense path is optional

Retrieval-over-markdown usually starts with a Python venv holding chromadb, plus a running Ollama with an embedding model pulled. That is a virtualenv, a background service, and a model download standing between install and the first answer.

• Default: BM25 lexical search, plus recency decay, plus a priority weight. Pure Python, zero external services, works immediately after install.
• recall index --dense opts into Chroma + Ollama for semantic retrieval on top of BM25.
• Everything that touches the dense path degrades instead of crashing when it isn't configured: retrieval fuses BM25-only, curation dedup falls back to BM25 lexical overlap, and recall index --dense without chromadb installed prints exactly what to run rather than a stack trace.

Install

pipx install atlas-recall

(or pip install atlas-recall inside a venv). To track the latest commit rather than the released version, install from source with pipx install git+https://github.com/Abydin/atlas-recall.

The dense retrieval path and the MCP server are both separate extras, not hard dependencies:

pip install "atlas-recall[dense]"   # Chroma + Ollama semantic retrieval
pip install "atlas-recall[mcp]"     # atlas-recall-mcp -- needs Python 3.10+

The base package (everything except the MCP server) stays on a Python 3.9 floor; the mcp SDK itself requires 3.10+.

The three commands, with real output

$ recall init ~/notes
Wrote config to ~/.config/atlas-recall/config.json
notes_dir = ~/notes

Next: recall install --client <claude-code|claude-desktop|cursor|windsurf|cline|codex>
      wires this into your AI client's config directly (merges, backs up, idempotent;
      pass --dry-run to preview first).

Then: recall index    (builds the search index)
      recall query "..."   (see what would be injected)

$ recall install --client claude-code
{
  "success": true,
  "path": "/Users/you/.claude/settings.json",
  "changed": true,
  "dry_run": false,
  "backup": null,
  "client": "claude-code"
}

$ recall index
index: discovered=3 changed=3 unchanged=0 removed=0 edges_verified=1 edges_divergent=0

$ recall query "what temperature should the water be for a light roast pour over"
== LIKELY-RELEVANT NOTES (read before acting) ==
water-temperature -- ~/notes/water-temperature.md -- hybrid retrieval (BM25+dense, RRF)
  BODY: 96C for light roasts, 92C for dark roasts. Off the boil by about 30 seconds.
coffee-brewing -- ~/notes/coffee-brewing.md -- pour-over ratio and grind size notes

That ran with no Chroma collection built and no Ollama process running - recall init never mentions either. recall hook (the piece that actually wires into Claude Code) reads a UserPromptSubmit payload on stdin and emits the same block as hookSpecificOutput.additionalContext; recall install --client claude-code writes that wiring for you. Prefer to do it by hand? recall init still prints the raw hook block as a fallback -- paste it into settings.json yourself, merging the "hooks" key if one already exists there.

The human path: search and trace the same corpus

$ recall find "chain lube"
bike-maintenance | doc | bike-maintenance | chain lube schedule | ~/notes/bike-maintenance.md

$ recall node coffee-brewing
id: coffee-brewing
...
-- 1-hop edges --
  link -> water-temperature [water-temperature]

$ recall trace coffee-brewing
coffee-brewing
  coffee-brewing -[link]-> water-temperature

$ recall map coffee
# coffee

## coffee-brewing  [coffee-brewing]
pour-over ratio and grind size notes
  -link-> water-temperature

$ recall verify
verify: checked=1 divergent=0

find is full-text search (SQLite FTS5, porter-stemmed). node shows one note plus its outbound [[wikilinks]]. trace follows those links recursively. map shows the shape of a topic: search hits plus what they connect to. verify reports links that point at a note that no longer exists, so drift is visible instead of silent.

recall index builds the SQLite FTS5 + wikilink graph that find / node / trace / map / verify read (and the Chroma collection, with --dense). It does NOT need to run before recall query / recall hook work: the BM25 corpus those read is built fresh, in-process, per query -- there's no separate BM25 index step, and that's a real strength, not a gap: the automatic path works the moment notes_dir is configured, no indexing required. One directory, one config file - index and query are still two interfaces on the same engine, not two tools that happen to sit in the same repo.

The MCP server

atlas-recall-mcp (needs pip install "atlas-recall[mcp]" and Python 3.10+) exposes the same engine as MCP tools over stdio, so a client calls them directly instead of shelling out to recall. Every tool returns structured JSON -- fields, not formatted prose -- so the calling model can act on pointers[i].score or edges[i].divergent rather than re-parsing English.

Tool	What it does
recall_query	Hybrid retrieval for a text -- the same ranking the Claude Code hook injects automatically.
recall_search	Full-text (FTS5/BM25) search, optionally filtered by doc_type.
recall_list_notes	List indexed notes, newest-modified first, optionally filtered by type.
recall_node	One note by id: full body plus its 1-hop edges.
recall_trace	Follow the wikilink/frontmatter edge graph outward from a note.
recall_map	The shape of a topic: search hits plus each hit's neighbors.
recall_verify	Report edges whose target note no longer exists.

recall install --client <claude-desktop|cursor|windsurf|cline|codex> wires this in directly; see "Supported clients" above for exactly which config file it writes.

Propose-only curation

Curation here is proposal-only: an agent proposes edits, a human approves every single one, and nothing is auto-written. That is a deliberate split from the usual design -- basic-memory, for instance, describes itself as bidirectional sync between users and AI assistants, so the agent writes notes as a matter of course. recall distill reads candidate notes (name, description, body, optionally which existing note to update) as JSON on stdin and returns a list of proposed ADD/UPDATE operations - an advisory near-duplicate warning on ADDs, an explicit, name-based opt-in required for UPDATE (never a fuzzy match deciding to overwrite something). It never writes a file.

$ echo '[{"description": "grind size for espresso", "body": "18g in, 36g out, 25-30s."}]' \
    | recall distill > proposed.json
$ recall distill-apply proposed.json
ADD 'grind-size-for-espresso' (type=reference)
  description: grind size for espresso
  body: 18g in, 36g out, 25-30s.
Apply this op? [y/N] y
  -> wrote ~/notes/grind-size-for-espresso.md

distill-apply is the only place in this package that writes to your notes, and it asks before every op, individually - there is no batch approve.

Configuration

recall init <dir> writes ~/.config/atlas-recall/config.json (or $ATLAS_RECALL_CONFIG). Everything the engine touches comes from that file or a CLI flag - nothing in this package points at a specific person's machine, corpus, or rules. Two fields worth knowing about:

• rules_path: an optional file (e.g. a house style guide) injected verbatim, undiluted, at the top of every hook block. Empty by default.
• keyword_rules: [[keywords], "rule line"] pairs - if a prompt contains one of the keywords, that one rule line gets echoed inline. Ships empty. See examples/keyword_rules.json for the shape; copy it, don't inherit it, the rules in there aren't yours.

Scoping memory per user or session

This package indexes one flat directory of notes -- there's no user_id or session_id concept inside a single corpus. The existing mechanism already gets you most of what that buys: point a separate notes_dir at a separate config (ATLAS_RECALL_CONFIG=/path/to/user-a.json recall ...) and you get a fully isolated corpus, index, and dense collection per user/session for free -- config.py's storage is already keyed per (config path, notes_dir) pair, so two scopes never collide on disk. Use one config per user/session rather than one shared notes_dir with an internal tag.

We looked at tagging memories within one shared corpus instead (closer to mem0/Letta's user_id field) and are deliberately not shipping it: doing it properly means touching the graph schema (knowledge.py derives a note's graph id from its filename/frontmatter name, globally, with no namespace -- two scopes reusing a name would collide as a duplicate id), the dedup path in distill.py, and every read/write surface (CLI, hook, MCP server). That's a real design change to a "one directory, one index" package, not an evening's addition, and a shallow version (filtering by a path prefix in Python without touching the graph or dedup) would be inconsistent enough to be worse than the honest answer above.

What's in the box

Path	What it is
bm25.py	Pure-Python lexical retrieval, the always-on default.
dense.py	Optional Chroma + Ollama semantic retrieval, guarded imports.
retrieval.py	RRF fusion, recency decay, priority tiebreak, admission floor.
knowledge.py	SQLite FTS5 + wikilink graph: find/node/trace/map/verify.
hook.py	The Claude Code UserPromptSubmit integration.
server.py / _mcp_entry.py	The MCP server (atlas-recall-mcp) and its optional-extra import guard.
install.py	recall install/uninstall: merge, backup, atomic write into a client's config.
distill.py / apply.py	Propose-only curation and its human-gated apply step.
eval.py / eval_fixture.py	recall eval: precision/hit measurement harness plus its synthetic fixture corpus.
warmer.py	Optional macOS fix for cloud-synced notes going "dataless".
cli.py	The recall console script.

Upgrading: rebuild the dense index

If you built a dense index with an older version of this package and then upgrade, run recall index --dense again before relying on semantic retrieval. Two changes since the initial release each independently make an existing on-disk dense collection stale:

• The dense metadata now carries a path-relative key per note (fixes same-filename notes in different subdirectories shadowing each other). An index built before this reads back as a warning on stderr and falls back to name-based matching, which can mis-resolve duplicate filenames.
• The Chroma directory is now keyed per corpus (notes_dir), so two separate note directories no longer share one collection.

Either change leaves a pre-upgrade dense index pointed at the wrong collection or missing the new metadata field. recall query degrades to BM25-only in that case rather than crashing, but you'll get better results by just re-running recall index --dense after upgrading.

Known limitation, carried over honestly

recall index (incremental) only re-derives outbound edges for notes it re-parses that pass. An edge whose TARGET note was deleted or renamed, but whose SOURCE note wasn't touched, doesn't get a chance to self-heal - it stays flagged divergent (recall verify will show it) until recall index --rebuild does a full pass. This isn't hidden: it's the tradeoff of incremental indexing being fast, and verify exists specifically so it's visible rather than silent.

License

MIT, see LICENSE.