cogvault

Get started · Why · Compared · How it works · Benchmark · Install · Quickstart · MCP · Fleets

Get started in 30 seconds

Claude Code plugin — MCP server plus a skill that tells the agent when to recall and when to record:

/plugin install cogvault --marketplace NBibikov/cogvault

Claude Code before 2.1.275: /plugin marketplace add NBibikov/cogvault, then /plugin install cogvault@cogvault.

Memory lives in ~/.cogvault/memory (set COGVAULT_TENANT to change it). The plugin adds /cogvault:remember.

Any MCP client, one line (needs uv):

claude mcp add cogvault -- uvx cogvault mcp --tenant ~/agent/memory

Add to claude_desktop_config.json, ~/.cursor/mcp.json, or your client's MCP config:

{
  "mcpServers": {
    "cogvault": {
      "command": "uvx",
      "args": ["cogvault", "mcp", "--tenant", "~/agent/memory"]
    }
  }
}

Already have Markdown memory? Point the tenant at it — nothing to import. Claude Code's auto-memory works as-is (frontmatter type, [[links]] and all):

M=~/.claude/projects/<project>/memory
uvx cogvault index  --tenant $M --ignore MEMORY.md   # first pass embeds, later passes are incremental
uvx cogvault search --tenant $M "how do we deploy"

The MCP server indexes on start by itself; the CLI search reads the existing index. --ignore MEMORY.md keeps the index file from competing with the cards it points to.

Unedited answers from a real session (Sonnet, plugin installed, a demo memory of six cards); only the waiting time is cut.

Related MCP server: auxly-memory-cli

Why

Most "AI agent memory" tools want to be an autonomous LLM daemon that summarizes your work into an opaque database or a graph you can't read. For a fleet of coding agents that just need to reliably recall a decision, a bug fix, or an infra detail, that's the wrong trade.

cogvault makes the opposite bet:

• Your Markdown files are the source of truth. Open them, edit them, git diff them. The SQLite index is a derived cache — delete it and it rebuilds from the files.

• One library, many tenants. Each agent gets an isolated memory namespace via its own directory. A process loads each embedding model once and shares it across every tenant it touches; with the stdio MCP server that means one small process per agent session, not a central daemon.

• No LLM in the loop. Ingest and retrieval are deterministic. Your agent is the LLM — it doesn't need a second one to remember.

• Local, private, offline. FastEmbed runs on-device. Nothing leaves your machine.

Compared to

Checked against each project's README and docs on 2026-10-05.

Pick something else when: you want an LLM to distil and merge facts for you (mem0), relationships between entities are the point (Graphiti), you want the agent to manage its own memory (Letta), or you want a richer notes app around the same Markdown idea, with Obsidian sync and a hosted option (basic-memory — the closest to cogvault).

Pick cogvault when: you run several agents and want each one's memory isolated in its own directory behind one process; you want recall to be deterministic and offline; and you want to measure it — cogvault eval scores recall on your agents' real queries and cogvault analyze lists what they tried to recall and couldn't.

How it works

Anatomy of a recall

Hybrid retrieval fuses semantic (vector) and keyword (BM25/FTS5) ranking with Reciprocal Rank Fusion, then applies optional temporal decay (recent memory outranks stale) and MMR (diverse top results, not five near-duplicates). Each card contributes only its best chunk, so one long file can't fill the whole result list.

Benchmark

Measured on real recall traffic, not synthetic questions: 66 queries sampled from the query logs of 6 live agent tenants (58% Ukrainian, the rest English), each judged against the actual cards — including answers that no configuration returned. One query has no answer in memory and counts as a gap, so 65 are scored. Every configuration was re-indexed from scratch on copies of the same tenants with cogvault 0.11.0.

What the numbers do and don't say:

• hit@1 is a tie. All four land within 0.54–0.57, and the 95% bootstrap intervals overlap almost completely. Real agent queries read like card titles, so the right card usually wins on its name alone.

• The gap is in the top 5. e5-small with the summary chunk puts the answer in the top 5 for 91% of queries vs. 80% for the default MiniLM and 77% for English-only bge on this mixed-language memory. That's what an agent reading 5 results actually feels.

• 65 queries is still a small sample. Treat differences under ~0.1 as noise. The aggregate numbers are in assets/benchmark.json; the queries are private and stay in each tenant.

Reranking

A cross-encoder reads the query and each candidate card together, so it can tell which card answers the question, not only which one is about the same topic. With rerank = "BAAI/bge-reranker-v2-m3-int8", cogvault scores the hybrid top 10 by the first 800 characters of each card file and reorders them. On the same 65 queries (cogvault 0.12.0, e5-small + summary chunk):

The reranker made 17 queries better and 3 worse. The 95% bootstrap interval of the hit@1 gain is +0.09 to +0.34. The cost is ~2 GB RAM and 1–2 s per recall on an Apple M-series CPU, so it runs only in the shared daemon and in cogvault eval (--no-rerank scores without it).

Run the same check on your own memory: put judged queries in <tenant>/.cogvault-golden.jsonl ({"query": "...", "relevant": ["file.md"]}, empty relevant = a known gap) and run cogvault eval --tenant DIR.

Choosing an embedding model

Agent memory is often not English-only. The default is multilingual so nothing is broken out of the box — but pick the model that matches your fleet's language mix (set COGVAULT_MODEL, or Config(model=...)). Switching models auto-rebuilds the index.

*From the benchmark above: 65 real queries over mixed EN/UK memory, cogvault 0.11.0. On English-only memory bge-small-en is a fine choice; on Ukrainian content it returns a negative relevance margin (a distractor outranks the answer), so it is unsafe for non-English memory. Run cogvault eval on your own vault to decide.

COGVAULT_MODEL=BAAI/bge-small-en-v1.5 cogvault index --tenant ~/agent/memory

Pin the model per tenant so it travels with the data instead of relying on every command exporting COGVAULT_MODEL (forget it once and a model mismatch silently re-embeds the whole index). Drop a .cogvault.toml at the tenant root:

# ~/agent/memory/.cogvault.toml
model = "BAAI/bge-small-en-v1.5"
# optional: recursive = true, strip_frontmatter = true, ignore_globs = ["Templates/*"]

Now cogvault search --tenant ~/agent/memory "…" uses the right model with no env var. Precedence: explicit --model / $COGVAULT_MODEL > .cogvault.toml > built-in default.

Install

uv tool install cogvault        # CLI on PATH
# or
pip install cogvault

Or skip installing and run it on demand with uvx cogvault …. The first run downloads the embedding model (~0.2–0.5 GB, once per machine).

Add it to Claude Code as an MCP server in one line:

claude mcp add cogvault -- uvx cogvault mcp --tenant ~/agent/memory

Also listed in the official MCP Registry as io.github.NBibikov/cogvault. Wheels are attached to each GitHub release.

Quickstart

# index a tenant's markdown memory
cogvault index --tenant ~/agent/memory

# search (hybrid semantic + keyword)
cogvault search --tenant ~/agent/memory "how do I restart the worker service"

# enable temporal decay (recent wins) and tune diversity
cogvault search --tenant ~/agent/memory "deployment steps" --half-life 30 --mmr 0.5

# only cards of one frontmatter type (user / feedback / project / reference / …)
cogvault search --tenant ~/agent/memory "hard rules for deploys" --type feedback

Memory cards

cogvault understands two lightweight Markdown conventions (both optional — plain files index fine):

• Frontmatter type — either flat (type: reference) or nested (metadata: → type: reference). Parsed at index time and filterable at search time (--type, MCP type param, search(card_type=...)). Every hit carries a type field; cards without frontmatter get null.

• [[wiki-links]] — link targets are indexed, and the top search result includes a related list of linked cards that exist in the index (ghost links are dropped; matching is by exact filename stem).

As an MCP server (Claude Code, Cursor, any MCP client)

claude mcp add cogvault -- uvx cogvault mcp --tenant ~/agent/memory

Exposes two tools:

• cogvault_recall — natural-language hybrid search over this agent's memory. Optional type param filters to one frontmatter card type; the top result includes a Related: line built from its [[wiki-links]].

• cogvault_record — save a fact; it's written as a Markdown card and indexed

Indexing a folder tree (Obsidian vaults, knowledge bases)

By default a tenant is one flat directory of .md files. For a nested vault (e.g. Obsidian, with 01-Projects/…, frontmatter, and folders to skip), opt in:

cogvault index --tenant ~/vault \
  --recursive \
  --strip-frontmatter \
  --ignore ".obsidian/*" --ignore ".trash/*" --ignore "Templates/*"

• --recursive walks subdirectories; files keep their path relative to the tenant, so two notes named Tasks.md in different folders never collide.

• --strip-frontmatter drops a leading YAML --- … --- block so its keys don't pollute the embedding.

• --ignore GLOB (repeatable) skips paths relative to the tenant root.

Same flags exist on search and mcp, and as Config(recursive=True, strip_frontmatter=True, ignore_globs=(...)) for the library. Indexing is incremental: the first pass embeds everything, later passes only re-embed changed files. (Reference: a ~3,500-note vault → ~9,500 chunks, first index ≈ 3–4 min, then warm recall in single-digit milliseconds.)

As a library

from cogvault import Vault, Config

vault = Vault("~/agent/memory", Config(half_life_days=30))
vault.reindex()
for hit in vault.search("where are credentials stored"):
    print(hit["score"], hit["file"], hit["snippet"])

Keeping a tenant healthy

cogvault doctor --tenant DIR reports what silently degrades recall: cards with no frontmatter or type, legacy timestamp filenames, frontmatter wrapped inside frontmatter, duplicate name: slugs, and [[links]] that resolve to nothing. Links resolve by frontmatter name:, filename stem, either separator style, and with or without the card-type prefix; links inside code and paths to files outside the tenant are not counted.

cogvault repair --tenant DIR fixes the mechanical half (dry run by default, --apply to write): unwraps nested frontmatter, infers a missing type, renames card-<timestamp>-….md to <type>_<slug>.md and rewrites every reference to it (MEMORY.md included), and adds minimal frontmatter to <type>_*.md cards that lack it. Healthy cards are left alone, and repaired cards keep their mtime so temporal decay is not reset.

Effectiveness logging

Every recall is logged (one JSONL line) so you can measure whether the memory is actually helping. cogvault analyze turns the log into a report:

cogvault analyze            # recalls, latency, per-tenant hit distance, weakest hits, card writes
cogvault analyze --json     # machine-readable
cogvault analyze --since 7  # only the last 7 days (or --since 2026-10-05)

The section that matters is weakest hits: each tenant's worst decile by vector distance — queries where recall returned something (hybrid search almost always does) but probably not the answer. That list is what your agents tried to recall and couldn't, i.e. the memory gaps to fill. Set COGVAULT_LOG=off to disable, or COGVAULT_LOG=/path.jsonl to relocate.

Multi-tenant fleets

Point one process at many tenants — each directory is an isolated namespace, proven by the test suite (test_multi_tenant_isolation). Agent B can never recall Agent A's memory unless you point B at A's directory.

Shared daemon

Each MCP client starts its own cogvault mcp. Without a daemon, each of these processes loads the embedding model (~1 GB), and none can afford the reranker. cogvault daemon loads each model once for all tenants and serves them over a Unix socket (~/.cache/cogvault/daemon.sock, mode 600). cogvault mcp and cogvault search find it automatically and forward to it; the MCP process then stays at ~40 MB. When the daemon is down, they search in-process as before, without the reranker, and log one daemon-fallback event to the query log.

cogvault daemon            # run in the foreground
cogvault daemon --status   # ping it: version, loaded models, tenants served

The daemon unloads a model after 10 minutes without a search or reindex, and loads it again on the next recall (~1.5 s extra on that call). Set --idle-unload SECONDS or $COGVAULT_IDLE_UNLOAD (0 keeps the models loaded). Measured with e5-small and the reranker: ~1.8 GB loaded, ~0.4 GB unloaded. ONNX Runtime uses 4 threads ($COGVAULT_THREADS, 0 = all cores) and no CPU memory arena, so a full reindex peaks at ~1 GB instead of keeping several GB after it. A shim that booted while the daemon was down waits up to 60 s for it before it loads a model itself, and drops a model it loaded as soon as the daemon answers again.

The reranker has its own thread count ($COGVAULT_RERANK_THREADS, default = the performance cores on Apple silicon, else 4). It runs only at recall time and is most of the recall latency: on 10 P-cores, recall p50 went from 2.3 s (4 threads) to 1.65 s, and to 1.05 s with onnxruntime 1.31 (that version also cut 4-thread recall to 1.5 s).

On macOS, run it under launchd with RunAtLoad and KeepAlive (program arguments cogvault daemon). After you upgrade cogvault, restart the daemon: launchctl kickstart -k gui/$(id -u)/<label>.

Design notes

Roadmap

• Importers (migrate existing memory from other stores)

• Pluggable embedders (Ollama, OpenAI-compatible endpoint)

• valid_until per-card temporal validity

• Optional FastMCP transport

License

MIT — your memory, your files, your infrastructure. Forever.

Figures are hand-built SVG from assets/make_graphics.py — python assets/make_graphics.py --png regenerates them.