droids-mem

README

droids-mem

An AI agent's memory, stored in a single SQLite file, zero external services.

Every time your agent learns something — a fix that worked, a project convention, a hard-won error resolution — droids-mem keeps it. Next session, that lesson is there when it's needed, retrieved by meaning not keywords.

One binary. One database. One brew install. No vector DB, no API key, no third-party service.

---

For whom?

You use an AI coding agent (Claude Code, Codex, OpenCode, Cursor, any MCP host) and you're tired of:

• Re-explaining the same project conventions every session.
• Watching your agent repeat mistakes it fixed last week.
• Wiring up vector databases, embedding pipelines, and API keys just to give an agent continuity.

droids-mem is local-first persistent memory that works out of the box. Your agent writes lessons; it reads the relevant ones at the start of the next run. That's it.

---

Quick start

# macOS / Linux
brew tap samuelmolero26/tap
brew install droids-mem

# Save your first lesson
droids-mem save \
  --task-type "my-project" \
  --kind error_resolution \
  --title "FTS5 deadlocks on trigger rebuild in same txn" \
  --what "ALTER TABLE inside a transaction, then rebuilding memories_fts in the same txn causes SQLITE_BUSY" \
  --learned "DROP the FTS table and INSERT-SELECT must run AFTER the parent ALTER commits"

# Search it later, even with different words
droids-mem search --query "fts trigger rebuild" --limit 5

# Browse everything interactively
droids-mem tui

All output is JSON on stdout; errors are JSON on stderr. Exit codes: 0 ok, 1 runtime, 2 usage, 3 not found, 5 conflict/duplicate, 10 dry-run.

Wire it to your agent

# One-shot: hooks, MCP bridge, and CLAUDE.md block
droids-mem install --all

# Or per host
droids-mem install --host opencode
droids-mem install --host codex

Done. Your agent now has persistent memory.

---

What it does

Memory

Agents save structured lessons — session summaries, task patterns, error resolutions, user rules. On save, droids-mem:

1. Scrubs secrets — API keys, tokens, credentials, PII. Detects 15+ patterns (AWS/GitHub/Stripe/OpenAI keys, JWTs, PEMs, emails, phones, private IPs). Rejects tags that match; redacts text fields.
2. Deduplicates — SHA-256 fingerprint on content + near-duplicate detection via BM25 + Jaccard similarity (≥ 0.85 threshold). Saving the same lesson twice is harmless.
3. Stores — one row in a local SQLite file.

On load, droids-mem assembles a two-tier context bundle for the agent: always-tier (last session summary + standing user rules, full body) and browse-tier (relevant error resolutions and task patterns, ranked by BM25).

Code graph (Go repos)

For Go projects, droids-mem builds a per-repo index of symbols and call edges (interface dispatch resolved, over-approximate). Instead of grep to find "what calls X", you get:

droids-mem graph symbol Store.Save --repo /path/to/project --direction up --depth 3
# → source + callers as signature stubs + transitive_callers count

Pre-built, signatures-first, agent-cheap. Auto-rebuilds on repo change; a repo that stops type-checking serves the last good graph flagged stale.

TUI

A three-pane terminal browser over the corpus — KINDS sidebar, a memory list, and a detail pane. Type to live-search (≥ 3 chars), tab cycles focus, ctrl+d deletes with confirmation. A CONNECTIONS view surfaces links between memories and their source files.

droids-mem tui

Session memory (Claude Code)

droids-mem hooks into Claude Code's lifecycle natively:

Event	What happens
SessionStart	Starts the MCP bridge; recovers crashed-run summaries
UserPromptSubmit	Injects relevant prior memories for the prompt
PostToolUse	Counts meaningful work (intake gate)
Stop	Once enough is unstaged, asks the model to record progress
SessionEnd	Saves the staged summary if the gate passes

Every hook fails open — a memory hiccup never breaks your session. Full reference: hooks/README.md.

---

Retrieval performance

The core bet: an agent should find a lesson later even when it's phrased differently than it was saved. This is measured, not asserted.

A fixed benchmark of 24 memories in seven distractor clusters (retrieval has to beat confusable neighbours) is queried by 33 hand-authored paraphrases whose wording is independent of the target. Runs in CI (internal/store/recall_benchmark_test.go); full report in eval/RESULTS.md.

Query class	recall@1	recall@5
Word-order / partial reword	100%	100%
Morphological ("cancel" → "cancelling")	100%	100%
Synonym, zero shared words	67%	75%
Overall	88%	91%

This is FTS5 + porter stemming — no embeddings, no vector DB. Pure-Go, no CGO. The honest ceiling is synonym substitution: a query that shares zero words with the memory (e.g. "too many requests" → a lesson titled "HTTP 429") can only be bridged by luck. Those misses are documented by name in the full report.

Reproduce:

go test ./internal/store -run TestRecallBenchmark -v

---

Install

Homebrew (macOS / Linux)

brew tap samuelmolero26/tap
brew install droids-mem

If Homebrew refuses the tap as untrusted: brew trust samuelmolero26/tap.

Go toolchain

go install github.com/samuelmolero26/droids-mem/cmd/droids-mem@latest

Requires Go 1.25+. Pure-Go (modernc.org/sqlite) — builds without CGO.

Prebuilt binary

Grab one for linux/{amd64,arm64} or darwin/{amd64,arm64} from the Releases page.

From source

git clone https://github.com/SamuelMolero26/droids-mem
cd droids-mem && go build ./cmd/droids-mem
./droids-mem --version

---

MCP tools

Six tools over the MCP bridge (bearer auth, or stdio for host-spawned servers):

Memory

• mem_save — persist a lesson (scrubs + dedupes).
• mem_search — full-text search (BM25 ranked).
• mem_context — two-tier context bundle for a task_type; mints a session_id.
• mem_get — fetch one memory by ID.

Code graph (Go repos)

• graph_symbol — a symbol's source plus callers/callees (and interface↔concrete implements edges) as signature stubs.
• graph_package — a package's exported surface, signatures only.

Graph responses render as TOON (Token-Oriented Object Notation) — one shared header per neighbor array instead of repeating JSON keys on every row — to keep "what calls X" answers cheap on hub symbols.

Operator commands (list, schema, doctor, migrate, prune, scrub) are not exposed over MCP — they're CLI-only by design.

droids-mem ensure-server   # ping /healthz, spawn detached serve if down
droids-mem serve           # foreground MCP bridge (Streamable HTTP)
droids-mem serve --stdio   # MCP over stdin/stdout (for codex, opencode, cursor)

Auth: Authorization: Bearer <token> on every /mcp request. /identity?nonce=<n> answers HMAC-SHA256(token, nonce) — anti port-squatting.

---

CLI reference

Command	What it does
save	Save a structured memory (scrubs + dedupes)
search	Full-text search across memories
context	Load a start-of-run context bundle for a task type
get	Fetch one memory by ID
list	List recent memories
tui	Interactive terminal browser
prune	Delete memories or find duplicate clusters
graph	Query a Go repo's code graph (index, symbol, package)
recent-sessions	List auto-saved session summaries
session	Session-memory plumbing (stage, check, flush, recover, hook)
install	Wire into a host: Claude Code hooks, or --host codex|opencode
uninstall	Reverse install: unwire hooks, deregister bridge, strip CLAUDE.md block
doctor	FTS integrity/rebuild, optimize, VACUUM, scrub stats
schema	Show parameter schema for a command
scrub	Run the scrub engine ad-hoc (--check, --test)
migrate	Establish the scrub baseline on an existing database
serve / ensure-server	Run or start the MCP bridge

Every command supports --help.

---

Shared context

Memory is local and private by default — every row has a scope (personal | shared) that defaults to personal, so nothing leaves your store implicitly. Opt in per memory from the TUI: flip rows to shared, then publish them into a dedicated git-tracked pool (shared.jsonl, one file per task_type) that teammates pull into their own store.

• Sharing is explicit and per-row — never a whole-corpus dump.
• Every shared copy is scrubbed at the trust boundary (local paths, session ids, and timestamps are stripped; content, tags, and kind cross over).
• Import dedupes across sources by fingerprint + Jaccard, so the same lesson from two teammates lands once.

Shared copies enter the git-tracked pool and can't be fully retracted — anyone who pulled keeps their copy. The TUI confirm dialog spells this out before push.

---

Secret scrub

Every save runs text through a single-pass scrub pipeline before it touches the database. Detectors run in three classes:

• Provider tokens: PEM keys, JWTs, AWS/GitHub/GitLab/Google/npm/Stripe/ Slack/Anthropic/OpenAI keys.
• Usage patterns: bearer headers, key = value assignments, URL credentials — gated on Shannon entropy so password = changeme survives but real secrets don't.
• PII: phone numbers, private IPv4 addresses, emails.

Longer redaction span wins on overlap. Tags that match any pattern reject the save (no silent auto-strip).

Field caps: title=200, what=8192, learned=4096, tags=500 — exceeding any returns field_too_large.

droids-mem scrub --check /path/to/some.log   # ad-hoc, no DB write
droids-mem scrub --test                       # run the fixture corpus
droids-mem doctor --scrub-stats               # aggregate counts across the DB

---

Configuration

All optional. Defaults match a single-user laptop install.

Var	Default	Notes
DROIDS_MEM_DB	~/.droids-mem/mem.db	Database file path
DROIDS_MEM_HOME	~/.droids-mem/	Token, pid, log files
DROIDS_MEM_MCP_TOKEN	Auto-generated	Bearer token for /mcp
DROIDS_MEM_MCP_ADDR	127.0.0.1:7777	Bind address (non-loopback logs a warning)
DROIDS_MEM_MCP_ENDPOINT	/mcp	/healthz + /identity always unauthenticated

State directory: mem.db (0600), token (0600), mcp.pid, mcp.log.

---

Troubleshooting

Boot gate on start. The database hasn't been through the scrub pipeline. The first non-bypassed command auto-runs migrate --rescrub (one-time write). If it fails, run droids-mem migrate --rescrub by hand.

db_init_failed. Check DROIDS_MEM_DB and that ~/.droids-mem/ is writable.

tag_contains_secret. A tag matched a scrub pattern. Tags aren't auto-stripped — fix the tag and retry.

scrub_emptied_learned. The learned field was fully redacted. Rewrite the lesson without the PII.

MCP bridge won't bind. Default 127.0.0.1:7777 conflicts if another listener holds the port. Verify using /identity?nonce=....

Stale FTS results. droids-mem doctor rebuilds memories_fts from memories if they diverge.

---

Contributing

Found a bug? Have an idea? Open an issue or pull request on GitHub.

This project is early and open to contributors who share its philosophy: local- first, no external dependencies, simple beats flexible. PRs that add new external services will be rejected.

---

License

MIT. See CHANGELOG.md for release history.