Hindsight Coding-Agent Memory

This machine runs the hindsight-coding-agents plugin: long-term project memory for coding sessions, backed by a Hindsight server. You (the agent) are already wired into it — this skill explains what happens automatically, which tools you have, and how to configure or debug it.

What happens automatically (no action needed)

• Per-repo memory bank: each repository resolves to a bank (shown in the session banner: ↳ memory bank “coding-agent::<repo>”). Worktrees share the main repo's bank.
• Ingestion builds itself: on first open, the bank is seeded from recent commit messages and a read-only codebase survey; every session start, a background engine tops it up (new commits, new conversations) and keeps 5 knowledge pages current. There is NO ingest command to run.
• Session synthesis: the first prompt of a session triggers one deep memory synthesis (reflect) injected into context. Later turns inject nothing automatically.
• Write-back: the session transcript is retained into the bank automatically at session end (per-turn on opencode). The user never needs to "save" a conversation.

Storing things deliberately

When the user says "store this in hindsight" / "remember this":

• The current conversation is captured automatically at session end — say so; no tool needed.
• An external document, notes, or durable findings → hindsight_ingest_document(title, content).
• A new feature/initiative being started → hindsight_capture_initiative(title, summary), right after the plan is agreed and before code is written.
• A plan that materially changed (goal, scope, or rationale — including mid-implementation) → call hindsight_capture_initiative again with relates_to_page_id set to that initiative's page id, summarising the current intent. Same page, updated plan — never a second page. Trivial course-corrections don't count.

Retrieving

• hindsight_search_knowledge_pages(query) — FIRST STOP for project questions (components, conventions, past decisions, initiatives). Server-side hybrid search, fast.
• hindsight_read_knowledge_page(page_id) / hindsight_list_knowledge_pages — read pages fully.
• hindsight_reflect(query) — deep reasoning over the whole memory for WHY questions and exact decided values; slower (seconds), use deliberately.
• Credit visibly whenever memory informs an answer: start that part with 🧠 From Hindsight memory (<page>): … — and never credit memory that didn't contribute.

Correcting wrong or stale memory

If you verify that something Hindsight served is wrong or outdated (the code, git, or an external source contradicts it), FIX THE RECORD — don't just ignore it. Call hindsight_ingest_document with:

• title: Correction: <topic> (e.g. Correction: retry policy 4xx set)
• content: (1) what memory claimed, (2) what is verifiably true now, (3) the evidence you checked (file/commit/output). Quote exact values verbatim.

Newer facts supersede older ones in retrieval, so one clear correction permanently outranks the stale memory. Do this whenever you catch a wrong injected memory, a stale knowledge-page claim, or an outdated decision — silent disregard leaves the trap armed for the next session.

Configuration — ONE file: ~/.hindsight/coding-agent.json

No environment variables (exceptions: HINDSIGHT_CONFIG relocates this file; HINDSIGHT_DIAG_FILE/HINDSIGHT_LOG_FILE/HINDSIGHT_LOG_LEVEL for diagnostics). Layering, later wins: defaults → file → harnesses.<name> → banks.<resolvedBankId>.

{
  "apiUrl": "http://localhost:8888", // your Hindsight server
  "apiToken": "…", // Hindsight Cloud only
  "gitIngest": "message", // "message" | "full" (per-commit diffs) | "none"
  "harnesses": { "claude-code": { "disabled": true } }, // per-agent override of anything
  "mapPathToBank": { "/Users/me/work/client-x": "client-x-memory" }, // path-prefix → bank
  "banks": {
    // per-repo control, keyed by RESOLVED bank id
    "coding-agent::secret": { "disabled": true }, // blacklist a repo
    "coding-agent::old": { "bank": "team::shared" }, // rename / converge banks (single hop)
    "coding-agent::mono": { "gitIngest": "full", "retainSessions": false },
  },
}

Key behavioral fields (any of them valid per-harness or per-bank): disabled, retainSessions (transcript write-back opt-out, history import included — recall and git ingest keep working), gitIngest, reflectTimeoutMs (AUTOMATIC session reflect, default 120000; hooks cap at 25s), reflectToolTimeoutMs/reflectBudget (the agent-invoked hindsight_reflect tool: default 330000 — above the server's 300s reflect wall timeout — and "high"), autoReflect (true; false = no injected first-prompt synthesis — the agent is instead told to call hindsight_reflect on new goals), pageRefreshEveryTurns (10), pageTriggerType/pageTriggerCron (when NEW knowledge pages refresh: auto-refresh (default) after each consolidation, cron on a schedule, manual never — existing pages keep the trigger they were created with), autoSeed/seedLimit (true/300), codebaseSurvey/surveyModel/surveyBudgetUsd (true/haiku/2), surveyRefreshCommits (0=off), logLevel ("info").

Config is read at process start, not watched: a hook harness picks an edit up on the next prompt, a persistent plugin (opencode, Kilo, Cline, Prime Agent, dsh) only after the agent restarts, and the MCP server in the next session. apiToken is the exception — re-read whenever the server rejects a request, so rotating it needs no restart. hindsight_diagnose reports the file's token and the running client's separately, which is how you tell a stale credential from a wrong one.

Blacklist a whole directory tree: map it to one bank and disable that bank — "mapPathToBank": {"~/scratch": "scratch"} + "banks": {"scratch": {"disabled": true}}.

Bank resolution order: mapPathToBank longest prefix → static bankId → template (default coding-agent::{gitProject}) → the matching banks.<id> section (its bank field renames the destination). Two repos share memory by converging their banks.<id>.bank on one name, or by one mapPathToBank prefix over their parent directory.

Install / update (for setting up another machine or harness)

npx @vectorize-io/hindsight-coding-agents install all     # every detected agent
npx @vectorize-io/hindsight-coding-agents install codex   # or specific: opencode|claude-code|codex|antigravity-cli|cursor-cli
npx @vectorize-io/hindsight-coding-agents uninstall       # removes exactly what install added
# updating is the same install command again — it re-copies the runtime in place

Debugging

• Readiness: hindsight_sync_status — synced: true = seeded memory queryable; also shows gitlog freshness, per-commit deepening progress, survey state (surveyDocs 0–4 = findings present; baseline without findings retries automatically), and active extraction ops.
• Logs: $TMPDIR/hindsight-coding-agent/plugin.log (leveled; set "logLevel": "debug" or HINDSIGHT_LOG_LEVEL=debug to mirror every event) and /tmp/hindsight-plugin.log (structured JSONL diag events with timings: session_start, reflect_ok, deepen_done, retain_ok, …).
• Reset a repo's memory: delete its bank on the server — the bank is the ONLY state; the next session is a true first-open. No client files to clean.
• Rule of thumb: memory silently missing → check the diag log for whether session_start/ deepen_started ever fired for that bank; a session started before the plugin was installed has no SessionStart behind it (its first prompt after install self-heals).
• Internal marker docs you may notice (safe to ignore, safe to delete): survey-baseline:<sha> — "🛰️ researching…" while a codebase survey runs, flipped to "✅ completed" once its findings land. Retained under the survey strategy, whose marker rule extracts NOTHING from status markers; powers the re-survey cadence and surveyBaseline in sync status. gitlog:<repo> is the aggregated commit-message seed document.
• Failures never break the agent: reflect/pages/retain failures degrade to a normal memoryless turn and are recorded in the logs.