lean-memory-console 0.2.1

Agent-first read-only verification console for lean-memory

lean-memory-console

A read-only verification console for lean-memory.

Agents write and search; the human verifies. Your agent (Claude Code via the observing MCP, or any HTTP client in Docker mode) is the only writer of memory content. You open the console read-only over stored memory — no adding, editing, or deleting facts — to see the engine's invisible signals for the first time: the ADD-only supersession spine, per-hit score decomposition, and provenance episodes.

The single write exception is the manual test-search box: it runs a real engine search and therefore bumps access stats (touch()). That is observability of live-search behavior, not a memory mutation.

Two modes

Local (default — zero Docker, nothing runs when the agent doesn't)

Two transient localhost processes over one shared data root:

# 1. The observing MCP wrapper — your agent stores + searches through this.
uvx lean-memory-console mcp

# 2. The console — you open this to verify. Binds 127.0.0.1, prints a
#    tokened URL, Ctrl-C to stop.
uvx lean-memory-console serve

serve accepts --root, --port, and --no-open (skip the browser launch). On load the SPA captures the token from ?token=… — held in memory and sent as the X-Console-Token header on every request — and strips it from the address bar (history.replaceState). There is no cookie; a page refresh without the token drops the session.

Via the Claude Code plugin (recommended), the MCP is wired automatically:

/plugin marketplace add Wuesteon/lean-memory
/plugin install lean-memory
/memory:ui          # launch + open the console
/memory:status      # resolved root, namespaces, connect snippets
/memory:server-up   # start the Docker data plane
/memory:server-down # stop it (the data volume persists)

(From a local checkout you can register the marketplace with /plugin marketplace add ./ at the repo root instead of the GitHub slug.)

Docker (single-tenant, long-running, container owns /data)

export LM_API_KEY=$(openssl rand -hex 24)     # REQUIRED — the container refuses to boot without it
docker compose -f "$(lean-memory-console --print-compose-path)" up -d

Then connect an agent over streamable-HTTP MCP and open the UI:

claude mcp add --transport http lean-memory http://127.0.0.1:8377/mcp \
  --header "Authorization: Bearer $LM_API_KEY"
# UI: http://127.0.0.1:8377/  (authenticate with the same LM_API_KEY)

Docker mode binds 0.0.0.0 so the published port is reachable; there is no local-mode Host guard and no per-launch session token. The controls are the bearer gate (LM_API_KEY) plus the MCP transport-security Host allowlist. If you reach the container over a LAN hostname/IP (no reverse proxy), list the host(s) in LM_MCP_ALLOWED_HOSTS (comma-separated; :* matches any port) or the /mcp mount 421s on its Host check — a commented example sits in the packaged docker-compose.yml. The compose service builds the full image target; container env is LM_DATA_ROOT / LM_API_KEY (required) / PORT / LM_CONSOLE_MODELS / LM_MCP_ALLOWED_HOSTS.

A plain REST mirror exists for non-MCP agents: POST /v1/{namespace}/memories and POST /v1/{namespace}/search.

The data-root rule (read this — it is the #1 trap)

The console serves exactly one data root and never auto-merges roots. Resolution order, both serve and mcp:

--root  >  $LM_DATA_ROOT  >  ~/.lean_memory

The ./lm_data trap. The core lean-memory engine's own default root is ./lm_data (a directory in your current working directory), not ~/.lean_memory. So if you ran the engine directly (or the core stdio MCP server pointed at the default) your memories may be under ./lm_data while the console defaults to ~/.lean_memory — and you would inspect an empty root.

• /memory:status prints the resolved root and warns when ./lm_data exists in the working directory but is not the served root.
• Fix: point the console at the right root, e.g. uvx lean-memory-console serve --root ./lm_data, or set export LM_DATA_ROOT=$(pwd)/lm_data.

The console adds exactly one file to the data root: _events.db (search/add traces). Everything else is the engine's own <namespace>.db files.

Live vs. replay: the t_ref rule

t_ref (epoch-ms) is the world/event time that becomes a fact's valid_at and anchors the temporal supersession spine.

• Live agents omit it. The observing MCP fills now, so facts are ordered by wall-clock ingest — correct for an agent capturing memories as they happen.
• Replay / import supplies it. When backfilling historical conversations, pass the original event time as t_ref (epoch-ms) on every add. Omitting t_ref on historical data silently collapses the spine's ordering — every backfilled fact gets now, so supersession and point-in-time queries become meaningless.

Rule of thumb: if you are importing anything that did not happen right now, set t_ref.

One namespace per project

Namespaces replace tenants: the engine stores one SQLite file per namespace, which is the isolation boundary. Use one namespace per project/session.

Cross-process writers on a single namespace (e.g. two Claude Code sessions spawning the wrapper on the same data root) are supported but serialized by a bounded retry-on-SQLITE_BUSY loop — not a lock manager. Pathological contention degrades to retries and, eventually, a SQLITE_BUSY surfaced in the event's payload.error. Keeping namespaces per-project avoids the contention entirely.

Namespaces are created implicitly on first accepted memory_add. There is no create-namespace step, and no delete-namespace surface (ADD-only discipline — to remove a namespace, delete its .db file while nothing is running). Namespaces whose sanitized name starts with _ are rejected (the _events.db sidecar is reserved); an empty name is not rejected — it sanitizes to default.

Image size (Docker full)

The default Docker image is the full target: it installs lean-memory[models] (CPU torch + sentence-transformers + a cross-encoder reranker), so the image is large (~1.46 GB) and the first run downloads model weights into the mounted HF cache volume. The CPU torch build is pinned from download.pytorch.org/whl/cpu so the resolve never pulls the multi-GB CUDA wheels. This is deliberate — stub vectors would recreate the FakeEmbedder first-impression failure the quality gate exists to fix. A slim target (~365 MB, stub embedder, no models) exists for API/UI development (docker build --target slim …) but is never the documented first-run path.

Offline by default

The console runs fully on deterministic stub backends with no network and no model downloads (LM_CONSOLE_MODELS=stub, or auto when [models] is not importable). When scores are stub-generated the UI shows a banner saying so.

License

Apache-2.0

Manual E2E verification (pre-merge)

Run this before merge. It is a manual gate, not automated — steps that need a live interactive Claude Code session are marked [manual] and are driven by a human, not by tooling. Capture the observed outcome next to each step.

Local mode (plugin + observing MCP):

• [manual] /plugin marketplace add ./ (from the repo root marketplace), then /plugin install lean-memory. Expected: plugin installs; the lean-memory MCP server appears in the session's MCP list, and /memory:ui|status|server-up|server-down are offered as commands.
• [manual] In a real Claude Code session, ask the agent to store a couple of facts and then search them (through the observing MCP memory_add/memory_search). Expected: memory_add returns {fact_ids, superseded_count}; memory_search returns hits whose fact_text matches what was stored; rows land in <root>/_events.db.
• [manual] Run /memory:ui. Expected: browser opens http://127.0.0.1:8377/?token=…; the address bar loses ?token after boot; no login screen (local mode).
• [manual] Verify all four pages render the resulting state: - Overview — the namespace card shows fact counts, top predicates, the 7-day adds/searches sparkline, supersession rate, and facts-per-add. - Memories — the fact table lists the stored facts; opening a fact drawer shows metadata, the supersession timeline (chain oldest→newest), and the provenance episode. - Episodes — the transcript shows the stored turns, each expanding to the facts extracted from it. - Activity & Traces — the polled feed shows the add/search events; a search row expands to the per-hit score decomposition (0.6·relevance + 0.2·recency + 0.2·importance, dense/sparse ranks, RRF); the test-query box runs a live search and its row is labeled origin: ui.

Docker mode (HTTP data plane):

• [manual] export LM_API_KEY=$(openssl rand -hex 24) then docker compose -f "$(lean-memory-console --print-compose-path)" up -d (or /memory:server-up). Expected: container starts; boot validation passes (data root writable, LM_API_KEY set, sqlite-vec loadable).
• [manual] claude mcp add --transport http lean-memory http://127.0.0.1:8377/mcp --header "Authorization: Bearer $LM_API_KEY". Expected: the MCP connection registers; a store+search from the agent succeeds and records events in the container's /data/_events.db.
• [manual] Open http://127.0.0.1:8377/. Expected: the login screen prompts for the key; entering LM_API_KEY authenticates; all four pages render the same as local mode, and the Overview shows the Docker connect snippet.
• [manual] /memory:server-down. Expected: container stops; the lm_data volume persists (re-up shows the same memories).