cambium
Cambium is a knowledge-lifecycle MCP server that captures, organizes, and promotes organizational knowledge across local, team, and org scopes, turning agents' work into reusable intelligence stored in git. Key capabilities: Capture knowledge manually via capture (memory, need, skill) and record_need. Automatically distill knowledge from agentsync claims and context-keeper decisions using distill. Import external knowledge from JSON/JSONL exports as provenance-tagged local items with import_memory. Recall knowledge across scopes with recall, which increments trust counters and abstains when uncertain. Promote knowledge through the scope ladder using promote, with endorsement gates, PR mode for org promotion, and generalization requirements. Endorse to fast-track local-to-team promotion and satisfy team-to-org requirements. Generalize project-specific org items into cross-project rules with generalize. Verify entries with verify_entry to update last_verified timestamps and keep promoted knowledge honest. Review promotions and generate stale reports using review_promotions and stale_report. Export markdown documentation (KNOWLEDGE.md) grouped by scope and project via export_markdown. Configure and monitor with setup (cold start configuration) and status (configuration gaps and metrics).
Provides knowledge storage and versioning using git repositories; knowledge items are stored in knowledge.json files in git branches with scoped access (local, team, org).
Supports org-scope knowledge promotion via pull requests on GitHub repositories, using the GitHub CLI for PR creation and management.
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:
distill()— turn events into memory automatically (passive capture)recall()— one federated read endpoint for every agent type (coding agent, Slack KB bot, SRE bot — same store, same call)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():
Scope | Lives in | Trust gate |
|
| none — it's yours |
|
| recalls ≥ N or an endorsement |
|
| an endorsement required; optionally lands as a pull request — review is the gate, |
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.
Related MCP server: lore
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:
env var | required | default | meaning |
| yes | — | your project clone (local scope, agentsync + context-keeper substrates) |
| yes | — | your unique agent id |
| no |
| git remote |
| no |
| per-invocation git/gh timeout (seconds) so a stuck call fails fast instead of hanging the server |
| no |
| team-scope branch |
| no |
| coordination branch name distill reads |
| no* | — | which repo holds the board — the same setting agentsync reads, so the two servers cannot disagree (see below) |
| no | — | path to the org knowledge repo clone (org scope off without it) |
| no | direct push |
|
| no |
| recalls needed for local→team |
| no | off |
|
| no |
| override the fallback config path (mainly for tests) |
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:
AGENTSYNC_BOARD_REPO— the explicit board address (env, or the same key in~/.cambium/config.json). One setting configures both servers.AGENTSYNC_REPO— agentsync's legacy explicit pin.CAMBIUM_REPO— but only if that repo actually holds the coordination branch (real ref lookup: local head → remote-tracking ref →ls-remote).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
statusbecomes"distilled_with_warnings"(not"distilled");top-level
warningscarries a plain-language line per skipped substrate;sources.agentsyncis 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
localscope. They have not earned promotion inside cambium and are not auto-promoted — team/org is still earned the normal way, throughrecall()usage andendorse().Provenance, always. Every imported item is stamped
source: {system, ref, imported: true, source_ts}and taggedimported, 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:
cambium field | source keys (first present wins) | if absent |
|
| record skipped (no body) |
— folded into body |
| omitted |
|
| empty |
|
|
|
|
| just |
|
| content-hash dedupe instead |
|
| omitted |
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 --openBuilds 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:
scope | items | recalled | recalls |
local | 229 | 15 (7%) | 117 |
team | 136 | 134 (99%) | 774 |
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 dashboard62 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 offorigin/mainand appends to the sameitemsarray inknowledge.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 againstknowledge.json.) Open a PR, merge it, then promote the next — serializing avoids the conflict entirely.Distill reads exactly ONE board per run.
distill()git fetchesclaims.jsonfrom<CAMBIUM_REMOTE>/<CAMBIUM_AGENTSYNC_BRANCH>of the repo resolved by board addressing —AGENTSYNC_BOARD_REPOif set, elseCAMBIUM_REPOwhen 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=1closes 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.
Available Tools
14 toolscaptureA
Save a knowledge item to your LOCAL scope: a fact, design note, gotcha, or troubleshooting step worth remembering. This is the manual capture path; distill() is the automatic one.
type : memory | need | skill kind : freeform subtype (note, decision, constraint, runbook, ...) why : the rationale — makes the item far more useful at recall time tags : comma/space-separated keywords (boost recall matching) valid_while : optional premise this knowledge depends on, e.g. "while we're on NetSuite" — surfaced later so a dead assumption is spottable
Inputs: content (required), why, kind, tags, +2 more. Returns: result
distillA
Automatically turn work that already happened into knowledge. Reads two substrates natively — no export step, no copy-paste:
agentsync: every DONE claim on the coordination branch (task + partner note + changed files) becomes an 'outcome' memory. The note your partner left for reconciliation is exactly the knowledge worth keeping.
context-keeper: every active decision and constraint in .context/ becomes a memory with its rationale, preserving the dec-NNN/con-NNN provenance.
Idempotent — each source record imports at most once; re-run freely (e.g. from a session-end or post-commit hook for passive capture).
Release-time capture (opt-in, CAMBIUM_RELEASE_CAPTURE=1): agentsync erases a claim from live state the moment it is released or re-claimed, so a claim that completes and churns before the next full distill is lost. With the flag on, distill also remembers the last-seen claim per agent and captures any that has churned away since the previous run — from that snapshot, via the same watermark, so nothing double-imports. Wire distill to fire on completion events and captured-once-at-completion is the result. The residual gap: a done state that lives and dies entirely between two runs is never observed (only the agentsync git log holds it).
Inputs: no named parameters. Returns: result
endorseB
Vouch for an item — the strong trust signal. One endorsement fast-tracks local->team promotion and is REQUIRED for team->org (usage alone never reaches org; someone has to deliberately say 'this is right').
Inputs: item_id (required), note. Returns: result
export_markdownA
Render knowledge to a human-readable KNOWLEDGE.md — grouped by scope then project, each item showing summary, kind, provenance (dec-NNN / claim origin), recall count and promoted date. cp1252 mojibake (em dashes, curly quotes) is normalized so the text is clean.
scope='org' (default): re-render the org knowledge repo's KNOWLEDGE.md from its knowledge.json and commit + push it beside the JSON, so the org repo's docs are always current. (This also runs automatically after any org promotion — direct-push commits both files together; PR mode puts both on the same PR branch.)
scope='local'|'team'|'all': render those scope(s) and RETURN the markdown without publishing — there is no repo to publish local/team docs to.
Inputs: scope. Returns: result
generalizeA
Restate an ALREADY-PROMOTED item's body as the cross-project rule, in
place, keeping the concrete version as example. The remediation
counterpart of the org generalization gate: items that reached org (or team)
before the gate — or were forced past it — are listed by review_promotions()
under org_needs_generalization; this rewrites one to its general form.
org_content : the cross-project rule to become the body. If omitted, the item's latest endorsement note is used (that is where the generalization was usually already written). note : optional note recorded as a verification stamp.
Writes through the org store's CAS path (direct push, or the shared
cambium/generalize PR branch when CAMBIUM_ORG_PR=1 — repeated calls batch
into one reviewable PR), re-rendering KNOWLEDGE.md alongside. Team-scope
items are edited via the team CAS path.
Inputs: item_id (required), note, org_content. Returns: result
import_memoryA
Ingest an external memory export into cambium as LOCAL-scope, provenance- tagged knowledge items — a source adapter alongside distill's substrate readers. Import/ingest only: it reads the source READ-ONLY and never writes back to it.
source : adapter name. 'json' = a generic JSON/JSONL export — a list of records (or an object wrapping one under memories/items/records), each with a text body (content/text/body/memory/note) plus optional title, why, kind, tags, id, timestamp. It's the extension point: new formats are new adapters, no core changes. path : local file path to read (no network, no external auth).
Every item is stamped with provenance (source.imported=True, the origin system, original id + timestamp) so imports never masquerade as native capture. Idempotent — re-importing the same records adds nothing (dedupe by source id, or content hash when no id). Imported items are NOT auto-promoted; they earn team/org the normal way, through recall usage and endorsement.
Returns a summary: imported / skipped / duplicates.
Inputs: path (required), source (required). Returns: result
promoteA
Graduate knowledge up a scope as it earns trust — the compound-growth step. With no arguments, scans your local items and promotes every one that qualifies to team. With an item_id, promotes that item one level (local->team, or team->org with to_scope="org").
Thresholds: local->team needs recalls >= CAMBIUM_PROMOTE_RECALLS or one endorsement; team->org always needs an endorsement (force=True overrides, use deliberately). Org promotion lands as a direct push, or as a pull request when CAMBIUM_ORG_PR=1 — the PR review is the org trust gate.
org_content : the cross-project restatement of a body that is specific to
one repo. Promotion to org changes the readership to everyone, so a
project-local runbook ("append to dashboard.py REGIMES") must become the
general rule ("annotate a regime boundary when a metric's computation
changes"). If the body reads project-specific and no org_content is given,
promotion is refused (with the tells and a suggested draft) unless
force=True. When supplied, org_content becomes the org body and the original
is preserved as example.
Inputs: force, item_id, to_scope, org_content. Returns: result
recallA
Search knowledge across scopes and return the best matches. THE read endpoint for every agent type — a coding agent, a Slack KB bot, an SRE bot — they all ask here, so knowledge captured once serves them all.
scope : auto (local+team+org, the default) | local | team | org limit : max results
Every returned item's recall counter is incremented (local directly, team best-effort via the shared branch) — usage is the trust signal promotion feeds on. If nothing clears the relevance floor the response says no_confident_match: true — don't present weak matches as established fact.
Inputs: query (required), limit, scope. Returns: result
record_needB
Record a NEED — something missing, wanted, or blocking (a first-class citizen alongside memories: 'we need staging seeds', 'docs for X are missing'). Needs surface in recall like any knowledge and can be promoted so the team/org sees recurring wants.
Inputs: content (required), why, tags. Returns: result
review_promotionsA
What's ready to move up? Lists local items eligible for team, team items eligible for org (endorsed), and org PRs already opened. The human-readable checkpoint before running promote().
Inputs: no named parameters. Returns: result
setupA
Finish cambium's setup in one call — the tool status() and every unconfigured error point you to. Validates paths, scaffolds .cambium/ (and gitignores it), and writes a local fallback config the server reads when env vars are absent (env still wins when set, and it takes effect immediately — no restart).
project_repo : absolute path to your project's git clone (required) agent_id : your unique agent id (required) org_repo : optional — a local clone path, OR a GitHub 'owner/name'. If a name isn'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. org_pr : optional — org promotion opens a pull request instead of a direct push. team_branch : optional — override the team-scope branch (default 'cambium').
No secrets are written anywhere; the config file holds only paths, ids, and flags, and lives outside any repo.
Inputs: agent_id (required), project_repo (required), org_pr, org_repo, +1 more. Returns: result
stale_reportA
Which promoted knowledge might be going stale? Lists team + org entries (the ones that cleared the trust gate) sorted OLDEST-VERIFIED-FIRST, with never-reverified entries flagged at the top and each entry's valid_while premise surfaced so a reader can spot dead assumptions ("while we're on NetSuite" long after the NetSuite migration).
project : limit to one project's entries (default: all) older_than_days : only entries last verified more than N days ago (plus every never-verified one); 0 = no age filter.
Event-driven, not clock-driven: this reports absent/old verification events, it does NOT compute a decaying confidence score. Re-confirm with verify_entry(); promotion also counts as a verification.
Inputs: project, older_than_days. Returns: result
statusA
First thing to call — especially when cambium looks broken. Returns structured config state: what's set, what's missing, what each gap costs in plain terms, and the exact setup() call that fixes it. NEVER raises on missing config. When fully configured it also reports item counts per scope/type, distill watermarks, and which substrates are actually wired.
Inputs: no named parameters. Returns: result
verify_entryA
Confirm a knowledge entry still holds — stamp its last_verified to now. This is the event that keeps promoted knowledge honest: promotion's trust gate defends what comes IN, verification keeps an entry from silently going stale after. An optional note records what was confirmed. Absent/old last_verified is a signal (see stale_report), never an automatic downgrade. Works on local and team entries; find stale ones with stale_report().
Inputs: item_id (required), note. Returns: result
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 14 tools
Each tool has a clearly distinct purpose: capture (manual save), distill (automatic import), endorse (vouch), export_markdown (render), generalize (restate as cross-project rule), import_memory (ingest external exports), promote (graduate scopes), recall (search), record_need (record a need), review_promotions (list eligible items), setup (initial config), stale_report (find stale entries), status (check config), verify_entry (confirm entry). Only minor overlap between capture and distill, but descriptions clarify the difference.
Tool names follow a consistent verb_noun pattern with underscores (e.g., record_need, review_promotions, export_markdown, verify_entry). Some are single verbs (capture, distill, endorse, generalize, promote, recall, setup, status), which is also consistent. No mixing of camelCase or other conventions.
14 tools is well-scoped for a knowledge management server covering capture, import, search, promotion, verification, reporting, and configuration. The number feels appropriate—not too few to miss functionality, not too many to be overwhelming.
The tool set covers the knowledge lifecycle comprehensively: creation (capture, distill, import_memory), search (recall), promotion (promote, review_promotions), endorsement (endorse), verification (verify_entry, stale_report), rendering (export_markdown), generalization (generalize), and setup. Missing explicit delete or update tools, but the domain may not require them (items can become stale). Minor gap.
Maintenance
Related MCP Connectors
Long-term memory for AI coding agents: durable project facts, recalled by every MCP client.
Persistent memory, knowledge graph, task management and more for AI agents via MCP.
Your memory, everywhere AI goes. Build knowledge once, access it via MCP anywhere.
Shared MCP memory for AI agents: knowledge base, skills, todos and agent-to-agent handoffs.
Related MCP Servers
- AlicenseAqualityBmaintenanceMCP-native, local-first memory for coding agents that turns real sessions into reusable decisions, gotchas, and domain knowledge.171,019 PyPI7MIT
- AlicenseNot gradedqualityBmaintenanceEnables AI agents to capture, store, and retrieve durable learnings from projects via MCP tools, providing a queryable memory of product and technical lessons across repos.MIT

sill-ensoulofficial
AlicenseAqualityCmaintenanceProvides persistent, cross-session memory for CLI agents via MCP tools, enabling them to recall, distill, and share experiences across projects and tools.86MIT- FlicenseNot gradedqualityCmaintenanceEnables AI coding agents and IDEs to capture conversations, extract structured knowledge into a queryable graph database, and recall past decisions, patterns, files, and errors across all projects and workspaces. Provides persistent memory tools for saving, searching, and cleaning up knowledge via MCP.2-