Yichen Agent Memory

Agent Memory Vault

Overview

Use this skill to help a user install or operate the public Agent Memory Vault system. The vault can be shared by Claude Code and Codex: Markdown remains the single fact source, while each host keeps a thin rule and hook adapter. Treat the GitHub repository as the product source, and treat this skill as the Agent-facing runbook for setup, maintenance, and troubleshooting.

Repository: https://github.com/mcncarl/agent-memory-vault

Safety Rules

• Never copy a user's real memory vault into a public repository.
• Never commit .env, SQLite databases, Zvec/vector stores, model caches, logs, API keys, cookies, tokens, passwords, private chat logs, customer data, or personal absolute paths.
• When migrating an existing memory system, migrate structure, scripts, and sanitized patterns only.
• Before publishing or pushing, run leak checks and inspect the diff.
• If a task involves deleting files, ask for explicit user permission first and use Trash, not permanent deletion.
• Never point Claude Code auto-memory at the formal vault. Either disable it or treat it as non-authoritative scratch memory.
• Shared cookies and tokens belong in a private config outside the vault and Git, with owner-only permissions; Agents should inject selected values into commands instead of printing the whole file.

Install Workflow

1. Clone or update the repository:

git clone https://github.com/mcncarl/agent-memory-vault.git
cd agent-memory-vault

2. Create a local private vault from the template:

python3 scripts/bootstrap.py --memory-root "$HOME/agent-memory-vault" --write-env
source .env

3. For a stable shared runtime, install the repository's canonical scripts and create a private TOML config:

python3 scripts/install_runtime.py --config-root "$HOME/.config/agent-memory"
cp config/agent-memory.example.toml "$HOME/.config/agent-memory/config/agent-memory.toml"
# Edit memory_root, git_root, and state_db. Never commit this private file.
"$HOME/.config/agent-memory/scripts/install_runtime.py" \
  --config-root "$HOME/.config/agent-memory" --verify --json

Treat the public repository as the only core-script source. The installer preserves unknown local adapters and private config while recording file hashes in runtime-manifest.json.

4. Optionally initialize Git for the private vault:

git -C "$AGENT_MEMORY_GIT_ROOT" init

5. Build indexes and validate:

python3 scripts/agent_memory_evolution.py --init --scan --report
python3 scripts/agent_memory_index.py --init --scan --report
python3 scripts/agent_memory_check.py
python3 scripts/agent_memory_doctor.py

6. Tell the user the final vault path, state database path, and the claim/closeout commands for future important tasks:

memoryctl --actor codex claim --file "/absolute/path/to/changed-memory.md"
memoryctl --actor codex closeout --dry-run
memoryctl --actor codex closeout

Daily Operations

When the neutral wrapper is installed, prefer it from both hosts:

memoryctl --actor codex search "query" --limit 5
memoryctl --actor claude prewrite "summary of the memory to write"
memoryctl --actor claude claim --file "/absolute/path/to/changed-memory.md"
memoryctl --actor claude closeout

Use only memoryctl and the agent_memory_* entrypoints. Compatibility wrappers are intentionally not installed; update every Hook, scheduler, shell alias, and custom script before removing an older installation.

Use the unified search entrypoint by default:

python3 scripts/agent_memory_search.py "query" --limit 5
python3 scripts/agent_memory_search.py "query" --track project
python3 scripts/agent_memory_search.py "query" --memory-type workflow

Before writing a new formal memory, run prewrite reconcile:

python3 scripts/agent_memory_closeout.py --prewrite "summary of the memory to write"

Allowed reconcile actions:

• ADD: create a new memory.
• UPDATE: update an existing memory.
• NOOP: do not write.
• MARK_OUTDATED: mark old information outdated without deleting it.
• MERGE_REQUIRED: stop and ask the user to merge or choose.
• ASK_USER: ask before sensitive, destructive, account, credential, cost, or uncertain actions.

Immediately after creating or changing a formal memory, claim it for the current host session:

memoryctl --actor codex claim --file "/absolute/path/to/changed-memory.md"
memoryctl --actor claude claim --file "/absolute/path/to/changed-memory.md"

Codex normally exposes CODEX_THREAD_ID. For Claude, configure agent_memory_session_hook.py --actor claude under SessionStart; it uses Claude Code's official CLAUDE_ENV_FILE mechanism to export the hook payload's real session_id to later Bash commands and clears any inherited Codex thread ID. Claims are stored in SQLite with only a one-way session hash. At the end of an important task, run closeout:

python3 scripts/agent_memory_closeout.py --dry-run
python3 scripts/agent_memory_closeout.py --commit

Session closeout processes only files claimed by that actor/session and also recovers their changes if an external backup tool committed first. Other sessions' files are excluded. A successful closeout records the processed content hash in memory_file_observations; only a matching observation proves that historical content is complete and lets the shared Git baseline advance. Unclaimed dirty memory must be resolved instead of being silently swept into a commit. Human maintenance can use memoryctl --actor human closeout --global deliberately.

Closeout uses both a process lock and a session ownership ledger: the lock serializes SQLite/Zvec/Git work, while claims prevent cross-session commits. Its compaction messages are non-blocking advisories, while failed checks or unresolved semantic duplicates still block the normal commit path.

Run audit manually when asked:

python3 scripts/agent_memory_audit.py
python3 scripts/agent_memory_audit.py --ignore FINDING_ID --note "reason"
python3 scripts/agent_memory_audit_autorun.py --reason manual --json
python3 scripts/agent_memory_doctor.py

Audit findings are prompts for review. Do not let audit directly rewrite Markdown facts unless the user asks for that update. Current-fact invariants can detect retired paths/scripts, wrong agent_scope, and stale fixed metrics in active summaries.

doctor is read-only by default. It also checks for stale session claims, an aging unpushed memory history, and a semantic virtual environment whose base Python disappeared. Use --repair-derived only when the user wants SQLite/FTS and Zvec rebuilt; it must never rewrite Markdown facts.

Optional Semantic Retrieval

Install the tested requirements-vector.lock, then build and validate Zvec:

python3 scripts/agent_memory_zvec_index.py --scan --prune
python3 scripts/agent_memory_zvec_index.py --report

For durable offline use, copy or APFS-clone a pinned model snapshot into the private runtime, set require_local_model = true, and record a private model manifest with revision, sizes, and hashes. doctor should pass the local-model, manifest-integrity, dependency-lock, and real offline-query checks before calling the semantic layer hardened.

The unified search runs SQLite and Zvec in parallel, applies all filters after merging, and discards semantic neighbors beyond the configured distance threshold. Search hits are candidates; always read the Markdown source before treating a claim as true.

Optional Automation

Only install global Codex hooks or a macOS LaunchAgent after the user has asked for automation.

• Stop is turn-scoped in both hosts, so the memory hook must be gated and quiet when the current session has no claims and all pending files belong to other sessions.
• Claude must have the SessionStart bridge in the live settings and in any settings manager's persistent configuration; verify it alongside Stop and SessionEnd after provider switches.
• A shared setup runs full closeout from Stop only after the Agent has written and claimed formal memory. Unclaimed dirty memory should block silent completion with a concrete claim instruction.
• Both hosts should block normal completion when closeout fails, using their native protocols: Claude returns decision: block; Codex exits with code 2 and writes a continuation prompt to stderr. Claude SessionEnd can be a non-blocking fallback; Codex currently has no direct equivalent.
• Keep the outer Stop hook timeout slightly above the closeout timeout. For a 300-second closeout, use at least 320 seconds outside.
• Keep one global closeout lock, one Git baseline, one session-claim table, one audit scheduler, one SQLite database, and one Zvec index across both hosts.
• Let agent_memory_audit_autorun.py --min-interval-days 7 decide whether audit is due.
• When the interval is due, autorun should run the content audit and then the read-only Doctor, persisting separate latest-audit.json and latest-doctor.json reports and notifying on either findings or infrastructure health drift.
• The weekly LaunchAgent must not use --force; otherwise closeout, hook, and launchd can run duplicate audits inside the same seven-day window.
• Merge the memory command into existing ~/.codex/hooks.json; never overwrite unrelated hooks.
• After changing a hook command, tell the user Codex may ask them to review/trust the new hook hash.

Upgrade Or Publish Workflow

When updating the public template repository:

1. Work in the agent-memory-vault repository.
2. Add or update only public-safe scripts, templates, docs, and fake examples.
3. Keep real user memory in the private vault.
4. Run:

python3 -m compileall scripts
python3 scripts/agent_memory_index.py --init --scan --report
python3 scripts/agent_memory_check.py --skip-state-db
python3 scripts/agent_memory_doctor.py
rg -n "/Users/|sk-[A-Za-z0-9]|token|secret|cookie|password|\\.sqlite|\\.db|zvec" .

5. Inspect git diff before committing.
6. Commit and push only after the leak check is clean or every match is confirmed to be a harmless example or warning.

Troubleshooting

• If search says the SQLite index is missing, run agent_memory_index.py --init --scan --report.
• If closeout cannot find changes, check AGENT_MEMORY_ROOT and AGENT_MEMORY_GIT_ROOT.
• If commit is skipped, read the warnings; MERGE_REQUIRED, ASK_USER, deleted files, or failed checks need user review.
• If Zvec is slow or unavailable, use --no-zvec for search/reconcile or --skip-zvec for closeout.
• If Zvec parity fails, run agent_memory_zvec_index.py --scan --prune and then --report.
• If audit repeats the same finding, record a decision with --ack, --ignore, --resolve, or --snooze; finding IDs must remain stable as counts change.
• If Claude debug reports zero matching hooks after installation, check whether a provider switcher rewrote ~/.claude/settings.json. Persist the hooks in that manager's common configuration and any live rollback copy, then restart it and verify the loaded matchers again.
• If doctor reports needs_review or mtime_fallback, do not fabricate a verification date. Classify structural/snapshot documents explicitly, use document dates only as provenance, and add verified_at only after checking real evidence.
• If doctor reports Zvec hash mismatch, run closeout for the claimed changed files or run agent_memory_zvec_index.py --scan --prune; equal document counts alone do not prove fresh vectors.
• If doctor reports session_claim_hygiene, preview with memoryctl --actor human claims-expire --older-than-hours 24 --json; add --apply only after confirming those sessions are no longer active. This changes the SQLite ownership ledger, not Markdown.
• If doctor reports memory_remote_backup, inspect the private-vault diff and leak scan before pushing. A clean local Git baseline is not a remote backup.
• If doctor reports semantic_python_runtime, recreate the private venv from the exact dependency lock with an available Python of the same supported minor version, then rerun the offline semantic probe.