Project memory
get_project_memory is the entry point for durable project memory. Its own description tells an agent exactly when to use it: “Call this before generating or repairing tests for an existing project; then call get_script only for the most relevant verified script to keep context and cost low.”
Parameters
Section titled “Parameters”| Parameter | Type | Required | Description |
|---|---|---|---|
project_id |
integer | Yes | Project whose memory should be recalled |
test_case_id |
integer | No | Filter to the most relevant verified runs for one case |
limit |
integer | No | Max verified scripts and MCP sessions returned (default 10, max 20) |
query |
string | No | Bounded task query used to recall relevant semantic learnings and agent episodes |
intent |
string | No | One of generate, repair, diagnose, inspect — only valid together with query |
downstream_reuse |
boolean | No | Client-reported outcome for a previously recalled memory, used only for aggregate telemetry |
What comes back
Section titled “What comes back”The response has a stable core plus additive reuse fields — older callers keep working unchanged as the contract grows.
Verified scripts — each entry carries script_id, framework, latest_passed_at, latest_execution_id, and a computed verification_status: current if the script hasn’t been edited since its last pass, or stale_after_edit if it has. An agent should re-run a stale_after_edit script before trusting it as verified.
Recent MCP sessions — prior negotiation sessions scoped to this project, with client_name, origin, tool_call_count, failed_tool_call_count, and the script IDs they touched.
Verification provenance — a plain-language sentence per verified script, for example “Last verified by Claude Code via mcp_stdio on 2026-07-09.” When no matching session exists it says so explicitly rather than guessing.
Project learnings — free-form notes attached to the project, plus a gated grounded_project_learnings view that only keeps lines anchored to something a verified run actually observed. See grounded learnings & evidence for how that gate works.
Built-in guidance
Section titled “Built-in guidance”The response embeds its own usage instructions rather than leaving an agent to infer them: recall_guidance says to reuse project_learnings first, call get_script for only the one matching script, and re-verify anything stale_after_edit. memory_reuse_guidance adds the same rule for the ranked-session and traversal paths — “Do not rediscover locators that held across passing runs.”
Recalling by query
Section titled “Recalling by query”Passing query (optionally with intent) additionally searches semantic learnings and past agent episodes — see recall & retrieval for how that search works and what its confidence fields mean.