Pact Memory

PACT Memory Skill

Persistent memory system for PACT framework agents. Store and retrieve context, goals, lessons learned, decisions, and entities across sessions with semantic search.

Overview

The PACT Memory skill provides:

• Rich Memory Objects: Store context, goals, tasks, lessons, decisions, and entities
• Semantic Search: Find relevant memories using natural language queries
• Graph-Enhanced Retrieval: Memories linked to files are boosted when working on related files
• Session Tracking: Automatic file tracking and session context
• Cross-Session Learning: Memories persist across sessions for cumulative knowledge

Store access: the two routes

STORE ACCESS. A memory operation (save, search, get, list, update or delete a record) goes through the pact-memory CLI. YOU DO NOT SELECT A STORE. Do not name a store by --db-path, by an environment variable, or by one more route somebody adds later. Let the CLI resolve it. A store you select is not the store the memory of the team lives in, so a save there is lost rather than shared. STORE INSPECTION is different: a row count, a column audit, or a schema check on the file. To inspect, do not run a CLI verb, do not import a module below skills/pact-memory/scripts/, and do not open the store read-write. In ONE command, against ONE resolved path, check that memory.db-wal and memory.db-shm are both absent by their full names, then open with mode=ro and immutable=1. Without immutable=1 the open fails. If a sidecar is present, stop and report. The read does not load the vector extension, so it cannot answer a question about vec_memories. Stop and report rather than take a barred route.

The sections below give the detail behind each sentence above.

MORE ON THE BRIGHT LINE, because an agent will look for an exception. THE RULE IS A POLICY, AND IT IS NOT A CLAIM ABOUT WHAT THE TOOL DOES. An agent does not select a store. The CLI can tell you to run setup --db-path <path>. That message is correct for the tool and it does not apply to you. You reach it only after you pass --db-path, which this rule forbids. Do not run it. Report it and ask the user. The danger is NOT that the flag reaches the live store. The flag binds the store you name, so the danger is that your save lands in a store nobody reads. For a memory system, a write that goes nowhere costs as much as a write that goes incorrectly.

To inspect the store

1. Check that memory.db-wal and memory.db-shm are both absent. Name the two files in full. Do not use a glob.
2. If one of the two files is present, stop. Report it. Do not read the store.
3. If neither file is present, open the store read-only: sqlite3.connect("file:<path>?mode=ro&immutable=1", uri=True)

What you must not do, and why

• Do not run a pact-memory CLI verb to inspect the store. The CLI opens a read-write connection and runs PRAGMA journal_mode=WAL on it, which creates memory.db-wal and memory.db-shm. A CLI read is thus a write to the store directory. A CLI run also reaches the same functions an import reaches, so it defeats the next rule.
• Do not import a module below skills/pact-memory/scripts/. In that package, functions create the live store directory as a side result of a path resolution, so an import puts each of them one call away. Do not read an absent import as a guarantee, because a subprocess reaches those functions with no import at all.
• Do not open the store with a plain sqlite3.connect. A read-write open creates the two sidecar files. That is a write to the store directory, even when no SQL runs.
• Do not copy the store with cp. A copy taken while commits sit in an uncheckpointed WAL is silently short of data, and it passes PRAGMA integrity_check. If you must have a copy, use VACUUM INTO.

Why the two flags, and why the check comes first

mode=ro alone cannot open this store. sqlite must create memory.db-shm to read a WAL database, and mode=ro forbids that creation, so the open fails with unable to open database file. immutable=1 tells sqlite to skip the WAL and read the main file directly, so the open succeeds and creates nothing. immutable=1 is not extra hardening. It is the only flag that opens the file.

The sidecar check is not a writer detector. It is a state selector. When the two sidecar files are absent, no WAL data is pending, so the main file is the whole database. That is the state in which immutable=1 is correct. If a sidecar is present, committed data can sit outside the main file, and immutable=1 then returns an older image and raises no error. One flag makes the read possible. The check makes the read correct.

If the check refuses

A refusal is information. In normal operation the CLI closes its connection and sqlite removes the two sidecar files on that close, so the check passes. A sidecar that stays is evidence that a writer did not close cleanly.

If a sidecar is present:

1. Report the sidecar by its full name and its size.
2. Do not read the store with immutable=1.
3. Do not delete a sidecar. A sidecar can hold committed data that is not in the main file at this time.
4. Ask the user before you go further. A recovery needs a write, and a write to this store is the user's decision.

To save a memory

Run the pact-memory CLI save command with no --db-path. The CLI resolves the default store. The section "Store access: the two routes" above gives the cause, and that cause covers a write.

Do not run python3 setup_memory.py init. That script runs as a script, it writes, and it accepts no --db-path, so the write reaches the live store and you cannot steer it away. The pytest refusal in the CLI is keyed on PYTEST_CURRENT_TEST, so it does not cover a run outside pytest.

Do not import a module below skills/pact-memory/scripts/ to write. Most of the modules in that package carry writing SQL.

Quick Start

All commands use the CLI entry point via ${CLAUDE_SKILL_DIR}:

# Save a memory
python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" save '{
  "context": "Implementing user authentication",
  "goal": "Add JWT refresh token support",
  "lessons_learned": [
    "Redis INCR is atomic - perfect for rate limiting",
    "Always validate refresh token rotation"
  ],
  "decisions": [
    {
      "decision": "Use Redis for token blacklist",
      "rationale": "Fast TTL support, distributed access"
    }
  ],
  "reasoning_chains": [
    "Redis chosen because TTL support → needed for token expiry → simpler than DB cleanup"
  ],
  "entities": [
    {"name": "AuthService", "type": "component"},
    {"name": "TokenManager", "type": "class"}
  ]
}'
# See Memory Structure table below for all available fields
# including agreements_reached and disagreements_resolved

# Search memories
python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" search "rate limiting tokens"

# Search with graph-enhanced boosting for current file
python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" search "auth tokens" --current-file src/auth/refresh.ts

# List recent memories
python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" list --limit 10

# Get a specific memory by full ID or unique prefix (>= 7 chars, case-insensitive)
python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" get <memory_id_or_prefix>

# Update an existing memory by full ID or unique prefix (scalar fields replace;
# list fields merge additively). Ambiguous prefix is refused.
python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" update <memory_id_or_prefix> '{"goal": "Updated goal"}'

# Delete a memory by full ID or unique prefix. Ambiguous prefix is refused.
python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" delete <memory_id_or_prefix>

# Check system status
python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" status

# Initialize/verify the memory system
python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" setup

All commands output JSON to stdout: {"ok": true, "result": ...}. Errors output JSON to stderr: {"ok": false, "error": "...", "message": "..."}.

For large JSON payloads (to avoid shell escaping issues), use --stdin:

echo '{"context": "...", "goal": "..."}' | python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" save --stdin

Memory Structure

Each memory can contain:

Task Format

{"task": "Implement token refresh", "status": "in_progress", "priority": "high"}

Decision Format

{
    "decision": "Use Redis for caching",
    "rationale": "Fast, supports TTL natively",
    "alternatives": ["Memcached", "In-memory LRU"]
}

Entity Format

{"name": "AuthService", "type": "component", "notes": "Handles all auth flows"}

CLI Reference

Commands

Exit Codes

Update Semantics

update uses additive merge with content-hash dedup for list-valued fields (lessons_learned, reasoning_chains, agreements_reached, disagreements_resolved, active_tasks, decisions, entities). Scalar fields (context, goal, etc.) still replace on update.

• Passing {"lessons_learned": ["new lesson"]} appends to the existing list; duplicate items (by content hash) are silently deduplicated, so repeated saves are idempotent.
• Pass --replace when you intentionally want to remove items from a list by overwriting it wholesale.
• Unknown top-level fields (e.g. {"foo": 1}) raise ValueError with exit code 2 instead of silently disappearing. Likewise, unknown sub-object keys (e.g. {"entities": [{"description": "…"}]} — the field is notes, not description) raise ValueError.

This prevents partial-list updates from silently clobbering the entire column.

Examples

# Save a memory
python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" save '{"context": "Bug fix", "lessons_learned": ["Check null values first"]}'

# Search memories
python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" search "authentication"

# Search with file context for graph-enhanced results
python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" search "auth patterns" --current-file src/auth/service.py

# List recent memories
python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" list --limit 5

# Update an existing memory — additive list merge (default)
# This APPENDS "New lesson" to the existing lessons_learned list; any
# existing lessons are preserved. Scalar fields like "goal" still replace.
python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" update abc123 \
  '{"goal": "Updated goal", "lessons_learned": ["New lesson"]}'

# Update with wholesale list replacement (--replace)
# Use this ONLY when you intentionally want to remove items from a list.
# After this call, lessons_learned contains exactly ["Only lesson that matters"]
# and nothing else.
python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" update abc123 \
  '{"lessons_learned": ["Only lesson that matters"]}' --replace

# Delete a memory
python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" delete abc123

Search Capabilities

Semantic Search

Uses embeddings to find semantically similar memories. Requires both:

• pysqlite3 — enables SQLite extension loading, which sqlite-vec needs
• model2vec — the embedding backend (model minishlab/potion-base-8M, 256 dimensions)

model2vec is the only embedding backend the code implements.

Graph-Enhanced Search

When searching while working on a file, memories linked to:

• The current file
• Files imported by/importing the current file
• Files modified in the same session

...are boosted in ranking.

Keyword Fallback

If embeddings are unavailable, falls back to substring matching across context, goal, lessons_learned, and decisions fields.

Setup

Dependencies

These are the three packages the memory system installs and imports:

# Enables SQLite extension loading — required before sqlite-vec can load
pip install pysqlite3

# Vector storage and similarity search
pip install sqlite-vec

# Embedding backend for semantic search
pip install model2vec

Install pysqlite3, not pysqlite3-binary. The two are different distributions and only pysqlite3 publishes artifacts for macOS on arm64.

setup installs all three automatically. Run it rather than installing by hand:

python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" setup

If any of the three is missing, semantic search is unavailable and search falls back to keyword mode. status reports which mode is active.

Initialize and Check Status

# Initialize the memory system (creates directories, database schema)
python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" setup

# Check system status (memory count, capabilities, db path)
python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" status

Storage

Memories are stored in ~/.claude/pact-memory/memory.db (matches a default-root pin in code — do not migrate) using SQLite with:

• WAL mode for crash safety
• Vector extensions for semantic search
• Graph tables for file relationships

Command Line Usage

When invoked via /pact-memory <command> "<args>":

Save Command

python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" save '<json>'

IMPORTANT: The argument is just a hint. You MUST construct a comprehensive memory object with ALL relevant fields. Never just save the raw string. Think of each memory as a detailed journal entry that your future self (or another agent) needs to fully understand what happened, why it mattered, and what was learned.

Required fields for every save:

Recommended fields:

• decisions: Key decisions made with full rationale, alternatives considered, and why they were rejected
• entities: Components, files, services, APIs involved (enables graph-based retrieval)

Writing comprehensive context:

BAD (too sparse):

"Debugging auth bug"

STILL BAD (single sentence):

"Debugging authentication failure in the login flow where users were getting 401 errors."

GOOD (comprehensive):

"Working on the fix/auth-refresh branch to resolve issue #234 where users reported intermittent 401 errors after being logged in for extended periods. The bug was reported by 3 enterprise customers last week and is blocking the v2.1 release. Initial investigation pointed to the token refresh mechanism, specifically a race condition between concurrent API requests. The authentication system uses JWT tokens with 15-minute expiry and a refresh token rotation pattern. This session focused on reproducing the bug locally by simulating high-latency conditions."

Example transformation:

# Agent is asked to save "figured out the auth bug"

# Construct the full memory object and save:
{
    "context": "Working on the fix/auth-refresh branch to resolve issue #234 where users reported intermittent 401 errors after being logged in for extended periods. The bug was reported by 3 enterprise customers last week and is blocking the v2.1 release. Initial investigation pointed to the token refresh mechanism, specifically a race condition between concurrent API requests. The authentication system uses JWT tokens with 15-minute expiry and a refresh token rotation pattern. This session focused on reproducing the bug locally by simulating high-latency conditions and tracing through the token refresh flow.",
    "goal": "Identify and fix the root cause of intermittent authentication failures that occur after extended user sessions, ensuring the fix doesn't introduce performance regressions.",
    "lessons_learned": [
        "The token refresh mechanism had a race condition: when multiple API requests detected an expired token simultaneously, each would trigger its own refresh, causing token rotation conflicts where subsequent requests used invalidated tokens",
        "Adding a mutex/lock around the token refresh operation prevents concurrent refresh attempts - the first request refreshes while others wait and then use the new token",
        "The bug only manifests under high latency conditions (>500ms API response time) because faster responses complete before the token expiry window, making it hard to reproduce in development",
        "Our existing retry logic actually made the problem worse by immediately retrying with the same stale token instead of waiting for the refresh to complete",
        "Integration tests should include latency simulation to catch timing-dependent bugs like this"
    ],
    "decisions": [
        {
            "decision": "Use mutex pattern for token refresh instead of request queuing",
            "rationale": "Simpler implementation with less state to manage. A mutex ensures only one refresh happens at a time while other requests wait. Our concurrency level (typically <10 concurrent requests) doesn't warrant the complexity of a full request queue.",
            "alternatives": ["Request queue with single refresh - more complex, better for high concurrency", "Optimistic token prefetch - would require predicting refresh timing", "Retry with backoff - doesn't solve the root cause, just masks it"]
        }
    ],
    "entities": [
        {"name": "AuthService", "type": "service", "notes": "Central authentication service handling login, logout, and token management"},
        {"name": "TokenManager", "type": "class", "notes": "Manages JWT token lifecycle including refresh logic"},
        {"name": "src/auth/refresh.ts", "type": "file", "notes": "Contains the token refresh implementation where the bug was fixed"}
    ]
}

Search Command

python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" search "<query>"

Returns semantically similar memories. Use natural language queries.

List Command

python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" list --limit 10

Shows recent memories (default: 20).

Sync Command

python3 "${CLAUDE_SKILL_DIR}/scripts/cli.py" sync

The Working Memory section of CLAUDE.md is a derived view of the store. sync replaces it with this project's newest 3 memories, each under its own date; to change what the section shows, update or delete the record, then run sync again. A sync_status of empty means the project has no memories and the file was not touched; check the envelope's project_id against the project you expected.

Best Practices

1. Save at Phase Completion: Save memories after completing PACT phases
2. Include Lessons: Always capture what worked and what didn't
3. Document Decisions: Record rationale and alternatives considered
4. Link Entities: Reference components for better graph connectivity
5. Search Before Acting: Check for relevant past context before starting work
6. Write Complete Sentences: Context should be a full description, not a fragment
7. Be Specific in Lessons: "X didn't work because Y" is better than "X didn't work"
8. Check Save Results: The save command verifies persistence by reading back the saved memory. If verification fails (exit code 2, error type SYSTEM_ERROR), the save silently failed — retry or check system status

Memory Layers: pact-memory vs Auto-Memory

The PACT framework operates with multiple memory layers. Understanding their distinct roles prevents duplication and ensures the right tool is used for the right purpose.

{config_dir} is this session's Claude config root — the value of $CLAUDE_CONFIG_DIR when set and non-empty, otherwise $HOME/.claude. Read it off an absolute path the platform already injected into your context — your plugin root is {config_dir}/plugins/… — rather than shelling out for the variable. Substitute it before running any command; never assume ~/.claude.

pact-memory's unique value: Structured fields (context, goal, decisions, lessons_learned, entities) enable semantic search, graph-enhanced retrieval, and cross-agent knowledge sharing -- capabilities that auto-memory's free-form markdown does not provide.

Coexistence model: Auto-memory captures broad session context automatically. pact-memory captures deliberate, structured knowledge at PACT phase boundaries. The Working Memory section in CLAUDE.md shows the most recent pact-memory entries, capped at 3, then reduced further by a token budget, so it can show fewer. The newest entry is not compressed and not dropped. Older entries are compressed to a one-line summary and keep their Memory ID. An entry carries no Memory ID when the record was saved without one, or when the identifier could not be written safely. What survives provides structured context that complements auto-memory's general learnings, and the full history stays searchable via the search command.

Integration with PACT

The memory skill integrates with PACT phases:

• Prepare: Search for relevant past context before starting
• Architect: Record design decisions with rationale
• Code: Save lessons learned during implementation
• Test: Document test strategies and findings

See references/memory-patterns.md for detailed usage patterns.