cambium

Part of the xylem stack.

The knowledge-lifecycle MCP that turns work your agents already did into compound, org-wide knowledge. cambium bridges two substrates that already exist — agentsync (what happened: claims, finishes, notes, changed files) and context-keeper (why: decisions, constraints) — and adds the three things neither has:

1. distill() — turn events into memory automatically (passive capture)
2. recall() — one federated read endpoint for every agent type (coding agent, Slack KB bot, SRE bot — same store, same call)
3. promote() — graduate knowledge local → team → org as it earns trust

Named for the cambium layer of a tree: the thin living tissue where all growth happens.

Why a third MCP (and not a database)

Knowledge layers fail because they're a side system nobody calls. cambium's bet: capture and recall must be native tools in the agent's loop, and state must live in the substrate the work already lives in — git — not a separate service. Storage is an implementation detail behind recall():

Team writes use the agentsync pattern: a private worktree under .git/ and git push as compare-and-swap, so concurrent agents never clobber each other.

Install

pip install -r requirements.txt      # just `mcp`

gh (GitHub CLI) is only needed for pull-request-mode org promotion.

Configure

Point an MCP client at the server — no env required to start:

{
  "mcpServers": {
    "cambium": {
      "command": "python3",
      "args": ["/abs/path/to/cambium_server.py"]
    }
  }
}

First contact is helpful, not cold. MCP servers can't start a conversation, so cambium teaches you through its own responses. Call status() (or any tool) before it's configured and instead of a bare env error you get structured guidance — what's set, what's missing, what each gap costs in plain terms, and the exact setup() call that fixes it:

{
  "configured": false,
  "gaps": [
    {"setting": "CAMBIUM_REPO",
     "cost": "no project repo → cambium has no substrate; every tool is unavailable",
     "fix": "setup(project_repo=\"/abs/path/to/your/clone\", agent_id=\"your-id\")"},
    {"setting": "CAMBIUM_ORG_REPO",
     "cost": "org scope off → promotions stop at team; org-wide recall unavailable",
     "fix": "setup(project_repo=…, agent_id=…, org_repo=\"owner/knowledge or /abs/path/to/clone\")"}
  ],
  "next_step": "setup(project_repo=\"/abs/path/to/your/clone\", agent_id=\"your-id\")"
}

setup(project_repo, agent_id, org_repo?, org_pr?, team_branch?) finishes the job: it validates the paths, scaffolds .cambium/ (and adds it to the repo's .gitignore), and writes a fallback config at ~/.cambium/config.json that the server reads when env vars are absent. It takes effect immediately — no restart. No secrets are written: the file holds only paths, ids, and flags, and lives outside any repo. If org_repo is a GitHub owner/name you haven't cloned locally, setup offers the exact gh/git commands to stand it up and leaves org scope off — it never creates or pushes a repo for you.

Env still wins. Any of the variables below, set in the MCP client config, overrides the config file per-key — the table is the full reference layer:

Where the agentsync board lives (board addressing)

* distill() reads finished agentsync claims off a coordination board. That board is a shared, long-lived team artifact — not a property of whichever project this session happens to be in — so cambium resolves its address independently of the session, using exactly the order agentsync itself uses:

1. AGENTSYNC_BOARD_REPO — the explicit board address (env, or the same key in ~/.cambium/config.json). One setting configures both servers.
2. AGENTSYNC_REPO — agentsync's legacy explicit pin.
3. CAMBIUM_REPO — but only if that repo actually holds the coordination branch (real ref lookup: local head → remote-tracking ref → ls-remote).
4. Otherwise: no board, reported loudly (next section) — never silently.

Why. Before this, cambium looked for the coordination branch in CAMBIUM_REPO while agentsync (unpinned) followed ~/.xylem/active_project.json. The two could point at different repos, and whenever the current project had never been provisioned neither found anything. distill() reported that as the bland string "no coordination branch found", callers treated it as normal, and the result was that distill imported zero agentsync claims across its entire lifetime — a three-legged design silently running on two legs.

A skipped source does not look like a completed one

distill()'s return now makes a miss impossible to read as a success:

• top-level status becomes "distilled_with_warnings" (not "distilled");
• top-level warnings carries a plain-language line per skipped substrate;
• sources.agentsync is an object — {status, board_repo, board_source, branch, claims_seen, done_claims, imported, reason, fix} — so "there is no board", "the board is here and nobody has finished anything", and "imported 3" are three visibly different results rather than one empty number.

status() reports the same under substrates.agentsync_board.

Org setup: create one (private) repo, e.g. github.com/you/knowledge, with an empty {"items": []} in knowledge.json; everyone who should read org knowledge clones it and points CAMBIUM_ORG_REPO (or setup(org_repo=…)) at their clone. cambium manages that clone (it hard-syncs it) — dedicate it, don't work in it.

Tools

capture(content, type, kind, why, tags, valid_while) — save a knowledge item to local scope (types: memory | need | skill). Manual path. valid_while optionally names the premise the item depends on, so a dead assumption is spottable later (see Machine-maintained documentation entropy).

record_need(content, why, tags) — first-class needs ("we're missing X"), promotable like anything else so recurring wants surface at team/org level.

distill() — the automatic path. Reads agentsync's coordination branch (every currently done claim: task + note + changed files → an outcome memory) and context-keeper's .context/ (active decisions & constraints, rationale and dec-NNN provenance preserved). Idempotent — wire it to a session-end or post-commit hook and capture becomes passive.

Release-time capture (opt-in, CAMBIUM_RELEASE_CAPTURE=1). agentsync keys claims by agent id and deletes a claim from live state the instant it is released or re-claimed — it exposes no hook or event, only the rewritten claims.json on the branch. So a claim that completes and then churns before a full distill runs against it is silently lost. With the flag on, each distill also remembers the last-seen claim per agent and captures any that has churned away since the previous run — reconstructing it from that snapshot, through the same dedupe watermark, so a claim captured at release time and again in a later full distill never double-imports. Fire distill() on completion events (a post-commit / session-end hook) and completed work is captured at its transition instead of only when a distill happens to catch it live.

What this is not: it is passive capture at the moments distill runs, not exhaustive reconstruction. The guarantee is precise — if a distill sweep observes a claim while it is done (or carries a note), that knowledge is captured even if the claim later churns. The residual gap: a done state that is created and churned away entirely between two sweeps (e.g. cambium wasn't running) is never observed, and only agentsync's git log still holds it. Walking that log to reconstruct such claims exhaustively is a possible follow-up (the history survives — agentsync's history() reads it), deliberately left out of this change.

import_memory(source, path) — ingest an external memory export into cambium as local-scope, provenance-tagged knowledge items (see Import below). Read-only against the source; imported items are not auto-promoted.

recall(query, scope, limit) — federated search across local+team+org. Every hit increments the item's recall counter (the trust signal promotion feeds on) and records cross-project use. Abstains honestly: below the relevance floor it returns no_confident_match: true instead of confident-looking noise. Each result carries endorsed_as — the item's endorsement notes surfaced as first-class context, since for a promoted item that is where its cross-project meaning was written.

endorse(item_id, note) — vouch for an item. Fast-tracks local→team; required for team→org.

promote(item_id, to_scope, force, org_content) — no args: scan-and-promote all eligible local items to team. With to_scope="org": push to the org repo, or open a PR when CAMBIUM_ORG_PR=1 (the team copy stays, annotated, until the PR merges). Promotion stamps last_verified — promotion is a verification. Generalization gate: org scope is read by every project, so a body that reads project-specific (names a file, a test_* id, a dec-/con-NNN ref, or its own origin project) is refused at the org boundary — restate it as the cross-project rule via org_content= (the concrete body is kept as example) or force=True to override. The refusal hands back the endorsement note as a ready draft. Mirrors the endorsement gate; the safe path is the easy path.

generalize(item_id, org_content, note) — the remediation counterpart of the gate: restate an already-promoted item's body as the cross-project rule in place, keeping the concrete version as example. For items that reached org before the gate (or were forced past it) — the ones review_promotions() lists under org_needs_generalization. Omit org_content to fall back to the item's latest endorsement note. Writes through the org CAS path (direct, or a single shared cambium/generalize PR branch when CAMBIUM_ORG_PR=1, so repeated calls batch into one reviewable PR); idempotent.

verify_entry(item_id, note) — confirm an entry still holds; stamps its last_verified to now (optional note). The event that keeps promoted knowledge from silently going stale (see Machine-maintained documentation entropy).

stale_report(project, older_than_days) — promoted (team + org) entries sorted oldest-verified-first, never-reverified ones flagged, each entry's valid_while premise surfaced. Reports staleness; never auto-downgrades.

review_promotions() — what's eligible for team, what's endorsed for org, which org PRs are pending, and org_needs_generalization — org items whose body still reads project-specific (crossed before the gate, or forced), each with the tells found and the endorsement note as a suggested restatement.

export_markdown(scope) — render knowledge to a human-readable KNOWLEDGE.md, grouped by scope then project (each item: summary, kind, provenance dec-NNN/claim origin, recall count, promoted date; cp1252 mojibake normalized). scope="org" (default) re-renders and pushes the org repo's KNOWLEDGE.md beside its knowledge.json; local/team/all return the markdown without publishing. It also runs automatically after any org promotion — direct-push commits both files together, PR mode puts both on the same PR branch — so the org repo's docs are always current.

setup(project_repo, agent_id, org_repo?, org_pr?, team_branch?) — finish configuration from a cold start (see Configure). Validates paths, scaffolds .cambium/, writes the fallback config; offers gh commands for an org repo rather than creating one. No secrets written.

status() — config state first: what's set, what's missing, each gap's cost and the setup() that fixes it (never raises when unconfigured). Once configured, also counts per scope/type, import watermarks, and wired substrates.

The compound-growth loop

stobie's agent finishes work ──agentsync──▶ done claim + note + files
                                                 │
jonny's cambium: distill()  ◀────────────────────┘
        │ outcome memory (local)
jonny + teammates: recall() ×N  ──▶ trust grows ──▶ promote() → team
        │ visible to every collaborator's agent
someone: endorse()  ──▶ promote(to_scope="org")  ──▶ org repo / PR
        │
ANY agent, ANY project, ANY type (SRE bot, KB bot): recall(scope="org")

Passive capture: add a Claude Code hook that runs distill at session end — capture then costs zero per-note effort. Run it on completion events too (a post-commit hook) with CAMBIUM_RELEASE_CAPTURE=1 and finished agentsync work is captured at the moment it completes, before a release or re-claim can erase it.

Import

import_memory(source, path) ingests knowledge from an external memory system into cambium. It is import/ingest only — it reads the external store and writes cambium items; it never writes back to the source (export is a separate, riskier feature and is deliberately out of scope). Import is modelled as a source adapter, the same shape as distill's substrate readers: an adapter reads records from a source location and yields normalized cambium knowledge items, which land through the same normalize-and-write/dedupe path distill uses — no second mechanism.

What import guarantees:

• Local scope, always. Imported items enter at local scope. They have not earned promotion inside cambium and are not auto-promoted — team/org is still earned the normal way, through recall() usage and endorse().
• Provenance, always. Every imported item is stamped source: {system, ref, imported: true, source_ts} and tagged imported, so imported knowledge is distinguishable from natively-distilled capture and auditable back to its origin (system + original id + original timestamp).
• Idempotent. Re-importing the same records adds nothing — dedupe is by the source record's stable id when present, else by a content hash, routed through the shared watermark path.
• Read-only against the source, and dependency-free (stdlib, local files only — no network, no external auth).

The json adapter (reference)

The one bundled adapter reads a generic JSON / JSONL memory export from a local file — no service-specific coupling. It accepts a top-level array, an object wrapping a list under memories/items/records/data/entries, or JSONL (one JSON object per line). Each record maps as:

type is always memory; malformed lines and records with no usable body are counted as skipped, never crash the import. The return value summarizes imported / skipped / duplicates.

# import_memory(source="json", path="/abs/path/to/export.jsonl")

Adapters are the extension point. json is the only format shipped — this is not a claim of support for any particular memory product. To ingest another system, add one adapter (a generator yielding normalized items) to IMPORT_ADAPTERS; core logic doesn't change. Adapters that require network access or credentials are intentionally not included here.

Machine-maintained documentation entropy

Trust-gated promotion defends knowledge on the way in: an entry only reaches team or org after it earns recalls or an endorsement. But nothing marked it going stale afterward. A fact that was true when it cleared the gate — "billing runs on NetSuite", "the staging DB caps at 90 connections" — stays trusted long after the premise dies. Worse, agents recall it, act on it, and cite it, so a wrong assumption doesn't just persist; it gets institutionalized, and the more it's used the more authoritative it looks. Promotion raises the stakes of being wrong without adding any way to notice you've become wrong.

cambium closes this with verification events and premise linkage, not confidence scores or time decay — both of which manufacture false precision. A 0.62-confidence memory implies a measurement nobody took, and "trust halves every 90 days" would quietly demote knowledge that is simply stable and correct. Instead every entry carries an optional last_verified timestamp (promotion counts as the first verification; verify_entry records later ones) and an optional valid_while premise naming the condition it depends on. Staleness is event-driven: stale_report sorts promoted entries oldest-verified-first and flags the never-reverified, and distill's release-time path surfaces the oldest-verified relevant entries right when work completes — so re-checking rides an existing workflow beat. Absent or old last_verified is a signal to a human, never an automatic downgrade. cambium reports the smell; a person decides.

Seeing it: tools/dashboard.py

python tools/dashboard.py --open

Builds a single self-contained HTML view of the whole mesh — every project's context-keeper log and the cambium layer distilled from them. Four views:

• Overview — the distillation funnel: recorded → active → distilled → recalled → promoted. Each stage a subset of the one above.
• Projects — sortable table of every store; click through to one project's full log, its supersession chains, its constraint scopes and its health flags.
• Knowledge — what distillation produced, what actually gets recalled, and where things sit on the local → team → org ladder.
• Health — mojibake, thin rationale, untagged, stale and never-recalled, each as a share of the population it is drawn from so a big store does not look unhealthy merely for being big.

It exists because counting was already possible and understanding was not. The first run answered a question no single tool could:

Promotion is what makes knowledge get read. Local is a staging area and its low recall rate is the system working, not failing — a conclusion only visible with both scopes in one view. The first version of this tool read .cambium/knowledge.json alone (local scope only) and reported "nothing has left local"; team knowledge lives on a git branch, and 136 promoted items were being recalled 774 times.

It also surfaced 84 items carrying cp1252 mojibake distilled before context-keeper fixed the transport — concentrated on the team branches that are recalled most. tools/repair_mojibake.py fixes the local stores; the team branches are shared git state and are left to a deliberate push.

Local by design. The output is gitignored. The mesh spans private repos and ones with no remote at all, so a published version could honestly show aggregates and nothing more; this one shows everything, because it is yours. Read-only, and test_dashboard.py asserts that byte-for-byte.

Test

python3 test_cambium.py      # server integration suite
python3 test_dashboard.py    # the mesh dashboard

62 cases against real git repos: markdown export (KNOWLEDGE.md grouped by scope then project with provenance/recalls/promoted-date, cp1252 mojibake normalized, auto-written alongside knowledge.json on org promotion in both direct-push and PR modes), write-time normalization (a substrate that feeds distill a cp1252-mangled em-dash lands clean in the canonical store, so recall() serves repaired text, not just the rendered .md), onboarding (unconfigured status() reports gaps with costs and fixes, every tool fails helpful when unconfigured, setup() configures from a cold start and its config takes effect in-process, env overrides the file, org names are offered as gh commands not created, non-git and missing paths rejected), capture/recall (+ honest abstention), distill from both substrates (exact agentsync claims format; exact context-keeper .context/ schema) with idempotency, post-promotion staleness (optional last_verified/valid_while fields, absent-field back-compat, verify_entry local + team round-trips, promotion stamps verification, stale_report oldest-first ordering + never-verified flag + age and project filters, release distill surfaces the verification prompt), release-time capture (off by default; a done claim survives a re-claim churn captured exactly once; a noted claim released before it reaches done is kept where a full distill would miss it), import (JSON + JSONL export → provenance-tagged local items, re-import dedupes, content-hash fallback without ids, malformed/missing fields skipped not crashed, imported items stay local and unpromoted, source left untouched), the full promotion lifecycle (recall-threshold, endorsement fast-track, org-requires-endorsement, PR-mode with gh stubbed), the org generalization gate (a project-specific body is refused at the org boundary with its tells and a suggested restatement; org_content= generalizes and preserves the concrete body as example; force=True overrides; a clean universal body is not over-blocked; recall surfaces endorsed_as; review_promotions self-reports org_needs_generalization; generalize() restates an already-promoted body in place, keeps the concrete as example, clears the flag, and is idempotent), a distill legacy-field fallback (a pre-v0.4 context-keeper decision carrying only rationale still distills its WHY), cross-project trust tracking, team-write CAS under a concurrent peer push, two real-agentsync integration tests (drive the actual agentsync claim / finish / release tools when the sibling repo is present — including the release-capture seam), and a real MCP stdio transport test. CI runs it on every push.

Limitations (honest ones)

• Lexical recall, not semantic. The scorer is token/substring overlap — deterministic and dependency-free. Swappable for embeddings later; the tool contract doesn't change.
• Org usage isn't tracked (recalls at org scope don't increment counters) — org items have already finished climbing.
• PR-mode promotion isn't transactional — the team copy stays (annotated with the PR URL) until a human merges. That's the point: review is the gate. A direct consequence: promote to org one item at a time in PR mode. Every promote(item_id, to_scope="org") branches off origin/main and appends to the same items array in knowledge.json, so two concurrently-open promotion PRs edit the same region of that file and the second can't merge cleanly until the first lands. (Observed: two knowledge PRs opened back-to-back on 2026-07-10; the second needed manual conflict resolution against knowledge.json.) Open a PR, merge it, then promote the next — serializing avoids the conflict entirely.
• Distill reads exactly ONE board per run. distill() git fetches claims.json from <CAMBIUM_REMOTE>/<CAMBIUM_AGENTSYNC_BRANCH> of the repo resolved by board addressing — AGENTSYNC_BOARD_REPO if set, else CAMBIUM_REPO when it genuinely holds the branch. It is no longer pinned to the project repo, but it still does not merge several boards: work coordinated on a different board, or through the agentsync-remote transport against a different backing store, is invisible to that run. Distill once per board.
• Distill captures at the moments it runs, not exhaustively. By default it imports agentsync's currently done claims; a claim released or re-claimed before any distill catches it live is lost. CAMBIUM_RELEASE_CAPTURE=1 closes most of that gap by capturing claims at their done/released transition (via a last-seen snapshot), but a done state created and churned away entirely between two sweeps is still only recoverable from agentsync's git log — a history walk that is not built here.

License

PolyForm Noncommercial License 1.0.0 — free for any noncommercial use.

Related

Part of the xylem stack.