| name | graphiti-usage |
| description | Discipline for writing to and reading from the shared Graphiti temporal knowledge graph. Use when adding cross-session facts, recalling decisions/incidents/vendor verdicts, or wiring per-project namespacing via group_id. Applies whenever the graphiti MCP server is reachable. |
Graphiti Knowledge Graph — usage discipline
Shared temporal knowledge graph reached via the graphiti MCP server (default http://host.docker.internal:8765/mcp/). One host instance backs all projects; namespacing is enforced per call via group_id — there is no automatic project isolation.
Scope model — three tiers
Every fact has a SCOPE. Pick at write time:
| Scope | group_id value | What goes here |
|---|
| fleet | the literal string "fleet" | TRUE cross-project methodology: tool-use rules, vendor verdicts that apply everywhere, organisation-wide policy. Should be valuable in an unrelated client project tomorrow. |
| host | the literal string "host" | Host-side agent operations: pb-graphiti plugin internals, ingest tuning (SEMAPHORE, MAX_EPISODE_CHARS), fleet-mgmt scripts, dotfiles, host config, agent-orchestration runbooks. Not relevant inside a project session. |
| project | the project's stable id (e.g. myapp, client-x-store) | Project-specific quirks: LIVE-equivalent branch, project-only test commands, module-vendor decisions that don't generalise, per-client preferences |
Resolve the project id deterministically:
- DDEV: read
$DDEV_PROJECT env var inside the container (set automatically).
- Generic:
basename $(git rev-parse --show-toplevel).
- Host with no project context (e.g.
~/workspace/ working with plugin internals or fleet-wide scripts): use host.
- Never use
default, main, session ids, or random suffixes.
Test for "is it fleet?": would you want this fact auto-surfaced when you open Claude in an unrelated client project six months from now? If yes → fleet. If "only when I'm working on the pb-graphiti plugin itself" → host. If "only in this one project" → project id.
Hard rules
-
Before EVERY in-conversation add_memory call: classify and confirm scope with the user.
Output ONE line in this exact form, then wait for confirmation:
Save to graph — scope: [fleet | host | project=<id>]. Reply g / h / p / correction.
g (or global / fleet) → use group_id="fleet". Apply the "is it fleet?" test above.h (or host) → use group_id="host". Plugin internals, fleet-mgmt scripts, host config.p (or project) → use the resolved project id.- Any other reply → treat as a correction; re-classify and re-confirm.
Default suggestion: project from a DDEV/git context; host from ~/workspace/ with no project; fleet only when the fact passes the cross-project test.
Exception: the PreCompact / TaskCompleted consolidation hooks shipped with this plugin run non-interactively (no user at the keyboard). They resolve scope deterministically from $DDEV_PROJECT → git toplevel basename → host, and write without prompting. Hard Rule 1 applies to add_memory calls Claude makes mid-conversation, NOT to those hooks.
-
From a project context, query ["<project-id>", "fleet"]. Surfaces fleet methodology in every project without leaking project A's quirks into project B AND without pulling host-ops noise into project work.
-
From the host shell (no project context), query ["host", "fleet"]. Surfaces host-side operational memory (plugin internals, ingest tuning, fleet-mgmt) alongside fleet methodology. Do NOT query ["fleet"] alone — that misses the host tier.
-
Never query across all projects without explicit instruction. A multi-project audit ("look across everything") is the only valid reason to omit the group_ids filter or enumerate every project's id.
-
Graphiti is for DOMAIN facts. Not for hard rules. Hard rules that MUST auto-load every session belong in CLAUDE.md / ~/.claude/projects/.../memory/. Graphiti is queried on-demand, not auto-loaded.
Fixing a wrong scope after the fact
If a node landed in the wrong group_id, fix it via the Neo4j browser at http://localhost:7474:
// Move ONE specific fact by name (preferred — avoids dragging unrelated nodes)
MATCH (n) WHERE n.group_id = '<wrong-id>' AND n.name CONTAINS '<keyword>'
SET n.group_id = '<right-id>'
RETURN n.name, n.group_id;
// Move the relationships attached to a moved node too
MATCH ()-[r]-(n {group_id: '<right-id>'}) WHERE r.group_id = '<wrong-id>'
SET r.group_id = '<right-id>';
// Move the episode that sourced the fact (so provenance stays consistent)
MATCH (ep:Episodic)-[:MENTIONS]->(n {group_id: '<right-id>'})
WHERE ep.group_id = '<wrong-id>'
SET ep.group_id = '<right-id>';
Verify before/after with MATCH (n) RETURN DISTINCT n.group_id, count(*) ORDER BY count(*) DESC.
When to WRITE to Graphiti
Write an episode when you learn ANY of:
- A decision with rationale ("we squash-merge to live because rollback is one
git revert")
- A non-obvious project quirk ("project X's LIVE-equivalent branch is
uat, not main")
- An incident root cause ("DISPLAY var lost during sweep on ; carrier moved to mounts.yaml")
- A vendor / module verdict ("Vendor Y blocked", "Module Z OK for this stack")
- A client preference ("invoices always DRAFT", "PR comments minimised as off-topic")
- A repeatable runbook step ("regenerate index after dependency update")
Don't write:
- Ephemeral session state, in-progress task lists
- Information directly readable from
git log / git blame / current file state
- Restatements of CLAUDE.md content
When to READ from Graphiti
Read at the start of any task that touches:
- A project area you haven't touched in this session (
search_nodes with the area name)
- A vendor / module / extension before recommending it
- A branching / deploy / merge step (check for project-specific rules)
- A decision that looks like it might already have a precedent
Cite the Graphiti episode UUID + summary when acting on a recalled fact, same as artefact citation in any investigation protocol.
Tool call shape
Adding an episode (AFTER scope confirmation per Hard Rule 1):
add_memory(
group_id="fleet", # or the resolved project id
name="<short title>",
episode_body="<the fact, with Why + How to apply>",
source="text",
source_description="claude-code-conversation://<session_id> [add_memory <YYYY-MM-DD>]"
)
Source-description tagging is mandatory for ad-hoc writes. Use the exact format above:
- Scheme
claude-code-conversation:// distinguishes ad-hoc in-session writes from PreCompact hook writes (which use claude-code-session://) and from bulk ingest (which uses file://, https://github.com/, mid:, etc.).
<session_id> is the current Claude Code session's id — available as $CLAUDE_SESSION_ID in shell context, or via the session metadata. If unavailable, use the literal string unknown.- The bracketed
[add_memory <date>] suffix tells operators when and how the fact was written, which is essential for auditing "what has Claude added to the graph this week" — see the README "Auditing automatic writes" section for the cypher query that surfaces these.
Searching from inside a project context (always pass BOTH project + fleet):
search_nodes(group_ids=["<project-id>", "fleet"], query="<your question>")
search_memory_facts(group_ids=["<project-id>", "fleet"], query="<relationship you need>")
Searching from a host shell (no project context):
search_nodes(group_ids=["fleet"], query="<your question>")
Citation discipline — surface the source
Every episode written via the bundled ingest scripts carries a source_description that's a real link or path back to the source content:
| Source | source_description format |
|---|
| Folder ingest (markdown/text/runbook) | file:///absolute/path/to/file.md |
Slack ingest (with --workspace-slug) | https://<workspace>.slack.com/archives/<channel-id> (<YYYY-MM-DD>) plus per-message permalinks inline in episode body |
| Slack ingest (no workspace slug) | slack:<channel>:<YYYY-MM-DD> (informational only — not clickable) |
| Email ingest | mid:<Message-ID> (RFC 2392 URI) |
| GitHub tickets ingest | https://github.com/<owner>/<repo>/issues/<n> (or /pull/<n>) |
| Magento modules ingest | file:///absolute/path/to/module-dir/ (trailing slash) |
| PreCompact consolidation hook | claude-code-session://<session_id> [precompact <YYYY-MM-DD>] |
| TaskCompleted consolidation hook | claude-code-session://<session_id> [task-completed <task-id> <YYYY-MM-DD>] |
Ad-hoc Claude add_memory in-conversation | claude-code-conversation://<session_id> [add_memory <YYYY-MM-DD>] (per the format above) |
Manual user add_memory | whatever the caller passed |
When you recall a fact and act on it, surface the source. Append the source_description in brackets after the claim — same discipline as artefact citation in the investigation protocol.
Example (good):
The fleet decided to ship the new checkout flow next Tuesday because the A/B test showed 4.2% lift at p<0.01 [src: file:///home/lucas/workspace/.../meeting-transcript.md].
Example (bad — claim with no source):
The fleet decided to ship next Tuesday because of the A/B test results.
To fetch fuller source content for a recalled entity, use get_episodes filtered by the entity's episodes field (returned in some search responses), or query by group_ids + a content keyword.
Pinned (always-loaded) facts
The plugin's SessionStart hook also loads any episodes written to group_id="initial_ingest" — capped at the most recent 20 — and surfaces them BEFORE the dynamic top-N recall. Use this group for facts you want Claude to see every session regardless of project context:
- Fleet-wide hard rules (caveman style, billing DRAFT-only, etc.)
- Universal client preferences
- North-star reminders
To pin a fact, call add_memory(group_id="initial_ingest", ...). To unpin, delete the episode via delete_episode. There is no auto-pruning — when the group exceeds 20 episodes, the SessionStart hook silently shows only the 20 most recent, so older pins effectively "fall off". Keep the pinned set curated.
Why this exists
Flat-file memory at ~/.claude/projects/.../memory/ scales to ~50 facts before the index becomes unreadable. Domain knowledge across a fleet of projects easily hits 500+. Graphiti handles supersession (X was true until time T, then Y) and cross-fact retrieval (entity-relation queries) natively — neither is possible in flat markdown.
