Knowl: Self-Retiring Persistent Memory for AI Agents

Your CLAUDE.md only grows. Knowl retires facts when they change.

Quick start · Why supersession · What gets stored · Features · Agent setup · Viewer · Requirements · Full reference →

Your agent starts every session blank, so you keep a CLAUDE.md. It only grows. Six months in it still names the database you migrated off last spring, and now the agent gets both answers.

Knowl is persistent memory for Claude Code, Cursor and Codex, over MCP or the CLI. When a fact is replaced, the old one is retired instead of competing with the new one. No API key needed. When Knowl isn't sure the new fact replaces the old, it leaves both active and hands you the knowl supersede command to say so.

Turn that off and retrieval drops from 98% to 47%. End to end, 90 to 73. How it was measured ↓

Forty seconds, one decision, three agents:

Quick start

Requires Node.js 22 or later. macOS, Linux and Windows.

npm install -g @dat999zx/knowl
cd your-project
knowl init

The published package is the same one in every case; each of these installs it and puts knowl on your PATH.

pnpm add -g @dat999zx/knowl
yarn global add @dat999zx/knowl
bun add -g @dat999zx/knowl

Or run it without installing:

npx @dat999zx/knowl init

Knowl runs on Node.js in all of these — Bun installs it, Node executes it. It bundles native addons (SQLite, tree-sitter, the embedding runtime), so running the CLI under the Bun or Deno runtime directly is not supported.

knowl init creates .knowl/, installs the project guidance files, updates .gitignore, and registers Knowl with whichever agents it detects. It also warms a local embedding model (~53 MB) in the background — init succeeds either way, and without it you still get keyword search.

That is the whole setup. You do not record memory by hand: your agent reads and writes it as it works.

Connecting an agent

knowl init registers the MCP server for every host it finds. Start a new session afterwards so the agent picks up its guidance, and it will query and write memory on its own.

gate means Knowl can refuse an edit that invalidates code another session is holding. Neovim and Kiro work the same way as Zed and JetBrains, through knowl acp. Cline needs one line pointing it at the shipped plugin. Hermes Agent gets a Python plugin, installed for you, that works in the terminal and in Hermes Desktop alike. Any other MCP client works with no integration at all.

Running agents in parallel? Every git worktree resolves to the main checkout's store — Conductor workspaces, Claude Code's isolation: "worktree", or your own scripts all share one memory, with nothing to configure. How that works, and its one limit →

→ Every host, and what each one can do · How agents use it · MCP tools and resources

The idea: memory that retires itself

Most memory systems are append-only. Storing "we moved to SQLite" leaves "we use PostgreSQL" active and retrievable, so the agent gets both and picks by rank. Knowl treats a same-subject write as a correction: the predecessor is marked superseded, drops out of normal retrieval, and stays queryable through knowl timeline.

That single behavior is most of the accuracy difference. On the MemoryAgentBench Conflict Resolution corpus — 455 facts, 100 questions about which fact is current, top-5 retrieval, no LLM reader:

Same corpus, same ranker, same query path. The only variable is whether the outdated fact is still active. This is a retrieval-level measurement in Knowl's own harness: it asks whether the current fact comes back first, with no model in the loop.

Verified end-to-end, in the benchmark's own harness

Because a number you score yourself is worth less than one somebody else scores, the same claim was re-run inside MemoryAgentBench's harness, scored by its own code, with an LLM reading what Knowl returned — the harder, fully end-to-end setup, at the largest context the task offers:

18,332 facts, 100 questions, substring exact match. Every row uses gpt-4o-mini as the reader, Knowl's included — the paper states it for all RAG and memory agents, so these are like-for-like. Knowl and agentmemory were measured here; every other figure is from the MemoryAgentBench paper, arXiv 2507.05257v4, Table 3. agentmemory is not evaluated in that paper — its published numbers are LongMemEval-S retrieval recall, a different task — so it was run through the same harness with the same config, and both adapters share one reader code path so neither can drift from the paper's own RAG handler. Method, mechanism and reproduction steps: FINDINGS.md.

Otherwise shown are every commercial memory system the paper evaluates, plus the highest scorer from each baseline family. The paper's table has changed between versions — BM25 read 56 in v1 and reads 48 in v4 — so the version is cited, not just the table.

Knowl's 90 was measured 2026-08-08 and independently reproduced at 89.0 on 2026-08-19 with the checked-in adapter; agentmemory's 79 is a single run. Every figure here is one run at temperature: 0.7, and the ablation gap moved 4 points between two runs of the same 6k cell, so read them to the point rather than the decimal.

Switching supersession off in that same harness drops Knowl to 73, and the gap holds across a 40× change in corpus size:

The two sections measure different things and are not comparable to each other: 98% is retrieval top-1 at 6K with no reader, 90 is end-to-end accuracy at 262K with one. Only the second is comparable to the published systems above. See benchmarks for the protocol, the checked-in results, and what the task does not cover — including multi-hop, where Knowl scores 7 against a 14-point retrieval ceiling.

Supersession is a correction, not a delete: the item, its assertions, and its history all survive.

Not a mock-up — the same sequence against the published CLI, recorded from demo.tape:

Sharing memory across a team: knowl.cloud

Everything above is local and needs no account. knowl.cloud is the optional hosted layer for when one machine is not enough:

• Shared workspaces. Knowledge written in one checkout reaches teammates' agents, with each repository still owning what it publishes.
• Browser agents. claude.ai and chatgpt.com cannot run a local process, so they connect over a remote MCP endpoint with a token scoped to one workspace.

Local-only remains a first-class way to run Knowl. Nothing here is required to use anything above.

What gets stored

Every atom has exactly one of seven categories:

Alongside the content, each atom keeps a status (active, deprecated, rejected, archived, superseded), a freshness flag, confidence, tags, source commit, affected paths, and optional evidence pointing at files, commits, tests, commands, URLs, or indexed code symbols. File and symbol evidence go stale on their own when the code moves, which is how an atom admits it may be out of date instead of asserting a version of the repository that no longer exists.

What Knowl deliberately does not store is your conversations. Lifecycle capture records bounded events and summaries — never prompts, transcripts, stdout, or environment variables. Raw transcript search exists as an opt-in, off-by-default index over files the host already wrote.

→ Knowledge model reference

How agents use it

knowl serve exposes the store over stdio MCP; knowl init registers it for you. The workflow the installed guidance asks agents to follow is short:

1. Query memory with the words that name the subject before reading repository files.
2. Use an active hit directly; inspect files only on a miss, conflict, or stale result.
3. Store durable findings, stated goals, and recurring diagnoses as you go, and correct contradicted memory rather than duplicating it.

In practice that looks like this — a new session, no context, nothing pasted in:

You     why did we pick SQLite over Postgres?

Agent   → knowl_query "sqlite postgres database choice"
        ← decision · Use SQLite · active · fresh
          "Keeps storage repository-local and simple to operate."
          alternatives: PostgreSQL, MongoDB
          tags: database, local-first

        SQLite keeps the store repository-local and simple to operate.
        Postgres and MongoDB were both considered and rejected on that
        basis.

The agent answered before opening a single file, and it knew the options you rejected — which the code cannot tell it, because rejected alternatives leave no trace in a codebase.

Full detail, and why each gap exists, in docs/hosts.md.

Where hooks are available, they own the session lifecycle: bootstrap context, capture, checkpoints, and finalization happen without the agent being asked. Where they are not, knowl task run, task start, task checkpoint, and task finish cover the same ground manually.

knowl init writes the MCP registration for every host it detects. To wire one by hand, the entry is the same everywhere:

{
  "mcpServers": {
    "knowl": { "command": "knowl", "args": ["serve"] }
  }
}

Use knowl.cmd as the command on Windows. Codex reads the same entry under mcp_servers.

→ MCP tools and resources · Lifecycle reference

What Knowl is for

Knowl does one job: keep a project's settled knowledge accurate for the agents working on it. Not user preferences, not chat history — the decisions, constraints, and architecture a project runs on, and which of them are still true today. Most stores sit in a codebase, and the drift and evidence tooling is aimed there, but nothing in the knowledge model requires one.

Three choices follow from that:

• Typed, not free text. A decision carries reasoning and the alternatives you rejected. A constraint is a rule that must keep holding. A state atom is expected to go out of date. Retrieval can rank on those differences; it cannot rank on paragraphs in a notes file.
• Governed, not append-only. Status, freshness, provenance, conflict identity, and supersession let the store tell you that something stopped being true. That is the whole difference between memory and an ever-growing pile of notes.
• Repository-local, not a service. The database sits beside the project it describes. No account, no egress, no vendor between you and your own project history.

Knowl is deliberately not a personalization layer. It has no opinion about your users, and it keeps no transcripts of its own.

Features

Everything below works from the CLI and from any MCP-connected agent, against the same local database. No account, no server, no API key. Each item links into the full reference for the detail — and for the limits.

♻️ Knowledge that corrects itself

Seven typed atom types, where a same-subject write retires its predecessor instead of sitting beside it. That one behavior is the 90-vs-73 difference. Evidence attached to a file or symbol goes stale by itself when the code moves.

conflicts · timeline · query --as-of · pr --since · index-code

🎯 Retrieval tuned for agents

Vector-primary with a bounded BM25 fallback, reranked by freshness, status, and confidence, so the current answer wins rather than the merely similar one. The embedding model is local and optional — without it you still get keyword retrieval, and nothing leaves the machine.

query · context --token-budget · config set-model · access

⏱️ Work that survives the session

On Claude Code, Codex, and Cursor, hooks own bootstrap, capture, checkpoints, and finalization without the agent being asked. A clean finish distills up to eight durable candidates. Park a workstream under a key and pick it up in any session, from any directory.

knowl posture maximal turns the watchful half on in one command — searching past sessions on a miss, flagging atoms whose files moved, and asking every so often what the session is relying on but never verified. All of it off until you ask.

task run · handoff · park · resume <key> · posture

🔗 Workspaces

Your API repo learned something the frontend repo needs. Link them and a query fans out, while each repository keeps its own database and its own ownership boundary. Open a shared peer atom in full by id, or finish that repo's work from here by naming it on the call. Knowledge a repo already holds is shared only when you promote it.

workspace init · workspace add · workspace promote --apply

📦 Reusable procedures

Package a procedure with its scripts under .knowl/skills/, then read it before it ever runs. Roll several atoms into one architecture summary deterministically, with no AI provider involved at all.

skill list · skill read · skill run · synthesize

💾 Your data, and getting it back

Checksummed JSONL export and import with four explicit policies for when the same atom changed in two places. Restore verifies schema, size, SHA-256, and SQLite integrity before touching anything, and takes a pre-restore snapshot first.

export · import --on-divergence · snapshot create · gc · doctor

🛰️ The sessions on this machine can see each other

Twenty agents across four repos, and none of them knew the others existed — so two hit the same failure and both start fixing it, and a third upgrades the engine the rest are standing on. Knowl records what each session is on, what it wrote this turn, and which failure it has claimed, then says so before the second session starts the same fix. Every host with Knowl hooks is in it and they see each other, Codex beside Claude Code. Prints nothing when you are the only one running.

fleet · knowl_fleet

The commands worth knowing on day one:

knowl query "auth design"              # search project memory
knowl list --unread                    # browse it — and see what nothing ever reads
knowl edit <item-id>                   # open one memory in the viewer to fix it
knowl state                            # the active memory, as a hierarchy
knowl conflicts                        # items that contradict each other
knowl timeline <item-id>               # every version an atom ever had
knowl context --token-budget 1500      # a fixed-size briefing for an agent
knowl pr --since origin/main           # knowledge your diff may invalidate
knowl fleet                            # every agent session live on this machine, and what it is on
knowl config list                      # every setting, its value, and how to change it
knowl doctor                           # setup, retrieval, and registration

• Seven atom types — listed above. Structure instead of one growing notes file.
• Automatic supersession — a same-subject write retires its predecessor. This is the 90-vs-73 difference above.
• Conflict identity — mark an atom exclusive and Knowl refuses a second active answer to the same question, instead of quietly holding both. knowl conflicts
• Full history — every version an atom ever had survives as an immutable assertion. knowl timeline <item-id>
• Time travel — ask what the project believed on a past date: knowl query "auth design" --as-of 2026-01-01T00:00:00Z
• Evidence — attach files, symbols, commits, tests, commands, or URLs to an atom. File and symbol evidence go stale by themselves when the code moves.
• Drift detection — knowl pr --since origin/main flags knowledge your diff may have invalidated, before you merge it, and knowl_drift asks the same question from inside the agent that wrote the branch. What it reports is a cited path that is gone, not one merely edited — that distinction is what keeps the signal readable.
• The claims drift cannot reach — drift watches files, and about half the store cites none. knowl status dates those instead, by how long since anyone last restated them, and names the ones furthest past their own category's cadence. It ranks rather than flags: for prose there is no evidence a claim became false, only the absence of anyone reaffirming it.
• Code intelligence — incremental Tree-sitter index over .ts / .tsx / .js / .jsx, so evidence can point at symbol:// locators, not just line numbers. knowl index-code
• Secret-safe writes — every write is screened for detected secrets, sensitive paths, and oversized content before it lands. Long-lived memory is the last place a credential should end up.

→ Knowledge model · Evidence and drift

• Vector-primary ranking with a bounded BM25 fallback, reranked by freshness, status, confidence, and recency — so the current answer wins, not merely the similar one. (This is the agent/MCP path; a single-repo knowl query from the CLI is lexical.)
• Runs offline. The embedding model is local and optional; without it you still get keyword retrieval. Retrieval never sends your query anywhere.
• Five bundled embedding presets, including a multilingual one covering 200+ languages, plus custom for your own ONNX model. knowl config set-model <model>
• Exact-identifier support — filenames, item IDs, and symbol:// locators still hit even when semantic similarity is weak.
• Token-budgeted context packs — hand an agent a fixed-size briefing with constraints pinned first, so non-negotiable rules never get truncated away: knowl context --query "auth rollout" --token-budget 1500
• Usage feedback — agents report whether a result helped, and knowl access shows what is heavily used, what is stale, and what keeps causing corrections.

→ Retrieval and context

• Automatic lifecycle on Claude Code, Codex, and Cursor — bootstrap, capture, checkpoints, and finalization happen through hooks without the agent being asked.
• Work loops for everything else — knowl task start, checkpoint, finish, or wrap a single command with knowl task run "Run tests" -- npm test.
• Promotion at session end — a clean finish distills up to eight durable candidates out of the session, and a command that has succeeded three times becomes a skill atom describing it.
• Handoff — leave one baton for the next session in this repo. It is delivered once, then archived.
• Resume keys — park a workstream under a short key you keep, and pick it up in any session, from any directory, any number of times later. knowl resume <key>
• Optional transcript search — off by default, and off means nothing exists on disk. Turn it on and past session prose becomes searchable, so a memory miss degrades to a slower lookup instead of amnesia. Keyword indexing keeps up on its own; semantic coverage is filled by knowl reindex --transcripts, because an embedding model does not belong in a per-turn hook.
• The recall gap — how often an agent edited a file this store already knew something about without ever retrieving it. Invisible from inside a session, because an agent that never retrieved an atom cannot notice the atom exists. Counted on every tool call, shown to nobody but you, in knowl status — and split between the main thread and subagents, because a subagent receives no prompt reminder and no server instructions, so its share is the only read you get on whether the bootstrap card alone carries the habit.
• The write gate's own score — with change impact on, the gate that would refuse an edit to code another session changed runs in shadow first, recording every refusal it withheld. knowl status prints the precision that produced, next to the bar it has to clear before it is allowed to block anything (≥95% over ≥40 adjudicated findings) — so the decision to arm it is made against a number instead of a hunch. Absent entirely until the gate has withheld something: a repo that never ran it has not scored 0%, it has measured nothing.

→ Tasks, sessions, and lifecycle

The other half of the same problem: not one session across time, but several at once. Claude Code keeps a registry of its live sessions and lets one message another; it records nothing about what any of them is doing, and no other host records anything at all.

• A roster at session start — who else is running, grouped by repo, own repo first. Empty when you are alone, so a single-session user never sees a line about any of this.
• Every host with Knowl hooks is in it, and they see each other. A Codex session appears on a Claude session's roster and the reverse. Liveness comes from the host's own session registry where it publishes one, and from recency where it does not.
• "Another session is already on this problem" — two sessions never see byte-identical output, so failures are matched on a normalised signature rather than raw text, and a claim is keyed to the problem rather than the file. The card names the peer, its files, and the exact call to make; a bare announcement of a conflicting edit is measurably no better than saying nothing.
• A pre-flight before a shared surface moves — hooks, host settings, migrations, lock files, and the knowl install every other session's hooks are running on. Advice on a channel the agent already receives, never a refusal.
• A stop-time nudge when this turn's writes invalidated a file another live session had read, joined through the read set rather than guessed. Shadow by default — it records what it would have said, because delivering it withholds a stop and that costs a turn.
• Only reachable peers are offered as something to message. A session on another host or under another config directory is listed and marked, and the card asks you instead — a card that told the agent to message a session it cannot address teaches it to skip the next one.
• Machine-level, not per-repo. ~/.knowl/fleet.db, beside the resume keys: a session in ~/work/api upgrading the engine is a fact ~/work/web needs. knowl fleet reads it from any terminal, inside a project or not.
• fleet.enabled ships on, and so do the cards — the roster costs a directory listing and says nothing when you are alone, and a card is advice on a channel the agent already reads. What ships quiet is what would cost you something: the per-turn digest, and the stop-time nudge that withholds a stop.

→ Who else is running

Your API repo learned something the frontend repo needs. Link them, and a query fans out — while each repository keeps its own database and its own ownership boundary.

knowl workspace init product      # create the workspace
knowl workspace add product       # run inside each repo that joins it
                                  # ...or --default-visibility repo to keep its writes private

knowl workspace promote                               # pick what to share from a list
knowl workspace promote --category decision --apply   # or name it outright

Joining a workspace shares what the repo writes from then on, and says so when it does; pass --default-visibility repo to decline. What the repo already knows is shared only when you promote it. Peer results are labeled with the repo that owns them, and a shared one can be opened in full by id — without its affectedPaths or evidence, which resolve against a checkout you are not standing in. A peer that is missing or unreadable is skipped and disclosed, never a reason for your local search to fail.

Writing into a sibling is deliberate rather than incidental. An agent names the repo on the call and that one call runs as that repo — its store, its config, its ownership rules, stamped as its own — exactly as cd-ing there has always behaved for the CLI. Name nothing and a foreign id is refused as before. Either way a repo's private knowledge stays private until it is promoted.

→ Workspaces

• File-backed skills — package a procedure with its scripts under .knowl/skills/, then inspect it before it ever runs. knowl skill list · read · run
• Deterministic synthesis — roll several atoms into one architecture summary with no AI provider involved: knowl synthesize --scope storage

→ Skills and synthesis

• Portable export/import — checksummed JSONL with four explicit divergence policies for when the same atom changed in two places. knowl export · knowl import --on-divergence newer
• Verified snapshots — knowl snapshot create writes a checksum manifest; restore verifies schema version, size, SHA-256, and SQLite integrity before touching anything, and takes a pre-restore snapshot first.
• Garbage collection that previews by default and protects anything recently used. knowl gc
• knowl doctor — one command that checks setup, config, integrity, schema, retrieval, vector coverage, agent registration, and workspace health.
• Optional AI — configure a provider for knowl ask and raw-text ingest. Every feature above works without one.

→ Portability and maintenance · Optional AI

See it: the local viewer

`knowl view