memshelf-mcp

Put your agent's memory on a shelf, hand it the index.

                              _          _  __
 _ __ ___   ___ _ __ ___  ___| |__   ___| |/ _|
| '_ ` _ \ / _ \ '_ ` _ \/ __| '_ \ / _ \ | |_
| | | | | |  __/ | | | | \__ \ | | |  __/ |  _|
|_| |_| |_|\___|_| |_| |_|___/_| |_|\___|_|_|
  ____________________________________________
 | INDEX >> | E-01 | E-02 | E-03 | E-04 | ... |
 |__________|______|______|______|______|_____|
        memory shelves for AI agents

What this is

Long-running agent sessions burn tokens re-sending history and lose detail to lossy auto-compaction. memshelf applies the docshelf pattern — tiny index in context, bodies fetched on demand — to the agent's own working memory:

1. Closed conversation topics, research dumps, and bulky tool output are offloaded to a local shelf as Markdown episodes.
2. Each episode carries an LLM-written, contract-validated digest that preserves decisions, rejected alternatives, artifacts, and open threads.
3. The agent keeps only INDEX.md (kilobytes) + digests in context and recalls exact sections via INDEX → episode → section slice over MCP.

Positioning in one sentence: claude-mem's loop, git's substrate, docshelf's navigation — episodic memory you can grep, diff, review, and carry between hosts. Private and local by default: the standard storage mode is a local git repo with no remote configured. The tool is public; the memory never is.

Measured, not promised

One week of dogfooding on the live shelf — full numbers and methodology in docs/demo.md:

Tokens are counted as chars/4 everywhere, so the ratios are estimator-independent; absolute counts move with the tokenizer.

Quick start

As an MCP server:

# Claude Code
claude mcp add memshelf -- uvx memshelf-mcp

// Claude Desktop (claude_desktop_config.json)
{
  "mcpServers": {
    "memshelf": { "command": "uvx", "args": ["memshelf-mcp"] }
  }
}

As a Claude Desktop extension — adapters/claude-desktop/: an .mcpb bundle installed from Settings → Extensions, with a Default shelf setting so calls need not repeat the path. Nothing has to be installed alongside it — not even Python.

As a Claude Code plugin — adapters/claude-code/: a /shelve skill plus SessionStart / SessionEnd / PreCompact hooks.

Or from the shell (pip install memshelf-mcp, Python ≥ 3.10) — the same loop, no MCP:

memshelf init   --shelf ~/my-shelf --name "My working memory"
memshelf shelve --shelf ~/my-shelf --slug 2026-07-23-topic --kind topic \
  --digest "What was decided, what was rejected and why, what stays open." \
  --section "Decisions=What was decided, and what was rejected instead — one line each."
memshelf recall --shelf ~/my-shelf --id 2026-07-23-topic --section Decisions --log
memshelf rebuild --shelf ~/my-shelf  # render ledger/INDEX from the episode (a bot's job on shelves that have one)
memshelf stats  --shelf ~/my-shelf   # claimed + realized savings
memshelf doctor --shelf ~/my-shelf   # exit 1 on integrity errors

pip install 'memshelf-mcp[semantic]' adds an embedding sidecar so search also finds paraphrases and the other language (memshelf semantic build --shelf ~/my-shelf once; MEMSHELF_SEMANTIC=off turns it off). Optional: the base install stays grep-only and as light as it is.

Tool surface

One verb per job; the same names over MCP (memshelf_*) and in the CLI (long-form descriptions: docs/tools.md):

The rules the tools enforce

The digest is a contract, not a convention. It is the only thing read at recall before fetching a body, so a weak one devalues the whole episode. lint_digest runs the same validator as shelve with no side effects (--strict turns warnings into failures); errors block a shelve, warnings do not — a pure reference digest legitimately carries no decision marker. A rejected digest is a feature: the tool prints exactly what to fix and writes nothing.

--amend re-runs the whole pipeline — redaction, the digest contract, composition — so an amended episode is exactly as guarded as a fresh one, which a hand-edit of the file never is. Amending a slug that is not on the shelf is an error, not a create.

The episode is the source; everything else is output. ledger.tsv, INDEX.md, stats.svg and each category's .meta.json are derived: shelve writes and commits the episode alone, rebuild renders the rest — delete all four and rebuild restores them byte-identically. That is what makes two sessions shelving in parallel a non-event: the merge is clean by construction. On a shared shelf, let a bot own the derived files on main — ready-to-copy workflows in adapters/shelf-repo/; rebuild --adopt migrates an older shelf once, rebuild --check is the CI guard.

Two consequences worth stating plainly, because getting them wrong costs a merge conflict:

• doctor reports no-ledger-row and stale-index immediately after a correct shelve — on every branch, main included. Nothing is broken: the episode is written, the derived files are not rendered yet. They clear on the next rebuild — the bot's run, on a shelf that has one.
• Do not rebuild and commit the derived files by hand to silence them. That is exactly the conflict class the split removes: a hand-regenerated ledger.tsv/INDEX.md/stats.svg meets the bot's, and the merge stops being clean by construction. Wait for the renderer; on a shelf without a bot, run memshelf rebuild --shelf . as its own step.

If those warnings persist for a day while episodes keep arriving, that is a different state — the renderer is not lagging, it is stopped — and doctor says so separately, as derived-stale at error severity. The day is counted from when the renderer could first see the work, not from the ledger's last commit: an episode pushed minutes ago onto a shelf whose ledger has not moved since yesterday says nothing about the renderer, and saying otherwise sent readers to a manual rebuild, which is the conflict this whole split exists to avoid.

That arrival is read from this clone's reflog for the tracked upstream — the one local record of when the ref moved here. A commit date is not a substitute: it says when the episode was written, and «shelve now, push when confirmed» is a documented way to work, so the two can be a working day apart. Where the reflog cannot say — a fresh clone starts an empty one, which is what CI and ephemeral agent sessions run in — doctor reports renderer-wait-unknown at the unknown level instead of picking a verdict: from there, «stopped» and «handed the work a minute ago» look the same, and the renderer has to be judged where it can be observed, on its own job's run. On a shelf with no upstream there is no renderer to be fair to, and the ledger's own age stays the clock.

The ref the renderer is judged by is the render branch, not whatever the checkout happens to track (#180). The bot renders one branch — the remote's default one, read as origin/HEAD when the clone has it, else origin/main, else origin/master — so a clone standing on a PR branch is not told derived-stale for an episode that sits on origin/<pr-branch> only. That episode waits for a merge, not for the renderer, and doctor says so with the warning upstream-not-rendered; one that did reach the render branch is judged there even from the PR branch, and a local commit on either stays episode-unpushed.

There was a third way to hold stale-index forever, and it is fixed rather than documented: docshelf split any episode past 50 KiB into section files beside it, shelve committed the episode alone, and from then on this working copy rendered an INDEX no other checkout could produce — no rebuild could clear it, and search answered with addresses that existed on one machine (#109). shelve no longer splits. A shelf that already carries such directories keeps reporting them as local-split-dir until memshelf prune-splits --shelf . --apply removes them; the episode file holds every section, so nothing is lost. It is a dry run without --apply, and a split directory that is committed is reported and left alone.

advise proposes, never writes. It answers the question the project was founded on — a dead topic has been occupying 30K tokens for forty minutes. The tool cannot see your window, so you tell it what is in there:

memshelf advise --shelf ~/my-shelf \
  --occupant 'auth refactor=42000,closed' \
  --occupant 'search dump=9000,idle=18' \
  --occupant 'Case B verdict=12000,live,episode=2026-07-22-case-b-verdict'

Three things keep it honest: it counts itself (INDEX + digests are in the report, not left out of it), it verifies episode= claims before proposing a drop, and it reports net — a topic too small to pay for its own digest is not proposed at all.

Rollup shrinks navigation and nothing else. When INDEX grows into a real share of your window, rollup collapses a period into one digest-of-digests and moves the originals to archive/ — still reachable by recall and search, every ledger row intact. The rollup digest is yours, not the tool's: synthesizing a quarter is the part a tool cannot do.

index-bloat is not what a rollup is for. INDEX lists your episodes, so its size grows with the shelf by design; its budget grows with the shelf too (INDEX_BASE_TOKENS + INDEX_TOKENS_PER_ENTRY × listed). Over budget therefore means entries are overpriced, never that there are too many of them — so doctor reports the cost of one line, and the fix is to trim it and rebuild. A rollup would remove entries and their allowance together and leave the price where it was. Having the two paired the other way is what made "archive a third of your memory" the standard way to silence a formatting problem.

purge deletes the working tree, not history. Retention is opt-in per episode (--retain-until); purge is a dry run until --apply — and even then git history still has the file. Real erasure is a deliberate filter-repo pass over the whole repository, never a side effect of a tool call, and the purge report says so.

resolve regenerates derived paths, never merges them — a derived file has no history, only a current correct value. The one file it unions is recall-log.tsv, because a recall is an event, not a fact about the episodes. Conflicting episodes are content, not mechanics: resolve reports them and steps aside.

A tool answer says when the code behind it is stale. The same code ships five ways, the version number does not move between releases, and twice a merged fix was not acting because the copy answering the calls was behind main (#125, #158) — found hours later, sideways, as what looked like a tool bug. doctor reports that gap as served-code-differs; since a doctor finding is read only by whoever runs doctor, the MCP server now says it in the response itself: when the served package's hash differs from the memshelf-mcp checkout next to the shelf (or $MEMSHELF_CHECKOUT), every envelope — success or error — opens with a warning carrying that code, both short hashes and where each lives. Identical code adds nothing. No checkout to compare with adds nothing per call either, but the initialize instructions say the freshness is unknown and how to make it known. MEMSHELF_FRESHNESS_WARNING=0 switches the per-call warning off where the difference is intended.

The design rationale behind each rule lives in docs/DECISIONS.md and docs/ARCHITECTURE.md.

One memory, multiple AIs

The memory is vendor-portable, and that is a measured fact, not a design intention: the same live shelf has been read and cross-written by Claude Code (Anthropic) and Gemini CLI (Google) through one shelf-spec server — protocol and field notes in docs/portability.md.

Status

M0 complete: the pattern was validated with zero code on a live shelf — retro-import of months of material, then a week of shelve-at-close (docs/M0.md). M1 shipped the server/CLI that enforces it, plus the Claude Code plugin. Next milestones with exit criteria: docs/ROADMAP.md; release history: CHANGELOG.md.

Documents

Rendered site: https://ignatenkofi.github.io/memshelf-mcp/ — including the week-report infographic from the dogfood shelf.

Project memory shelf

shelf/ is this project's own memory shelf, dogfooding the tool it ships: session episodes with decisions, rejected options and their reasons. Before asserting anything about past work here, read shelf/INDEX.md and fetch the one episode it points at — a guess about a past decision is a defect, not an estimate. Derived files (ledger.tsv, INDEX.md, stats.svg, .meta.json) are written by the same commit as the episode (memshelf rebuild --shelf shelf); the shelf-pr-guard workflow fails a PR or a push to main where they drifted.

Origin

Designed as RFC-0001 in the docshelf-mcp repo (#42, #43, #44); this repo is the project's home from 2026-07-13 on. The docshelf copy is frozen as a historical snapshot.

Related projects

• docshelf-mcp — the sibling project and storage layer: PDFs/Markdown → chat-project-friendly document shelves with the same index-and-fetch economics (measured: ~3.7K tokens vs 1.2M per question). memshelf was born as RFC-0001 in its repo and reuses its splitter/indexer/read/search verbatim.
• The dogfood memory shelf is a private repo — by design (MANIFEST principle 5).

License

MIT — see LICENSE.

mcp-name: io.github.ignatenkofi/memshelf-mcp