basemem memory

Agents lose context between sessions, forcing you to re-explain architectural decisions and project constraints. This skill provides a local-first persistent memory layer that stores decisions, project state, and code-graph evidence in a SQLite database.

What it does

• Persistent memory stores decisions, facts, and architectural constraints across multiple agent sessions.
• Code intelligence indexes project symbols, callers, and callees to provide structural context without full-file reads.
• Session management tracks progress with session starts, recaps, and end summaries to maintain continuity.
• Conflict resolution identifies and manages contradictory notes or outdated project information.
• Local storage maintains all data in a local SQLite database with no requirement for external vector stores or hosted services.

How it works

1. Initialize the memory by calling code_init to index the project and update_planet to set the initial goal.
2. Record significant decisions mid-work using logInteraction to ensure constraints are preserved.
3. Recover context at the start of new sessions using session_recap or read_planet.
4. Search for specific historical context using search_notes or code_find when encountering unfamiliar symbols.

Frameworks & tools

This tool requires Python 3.10+, Node.js, and Git. It utilizes a local SQLite database for storage and optionally integrates with ripgrep for fast text searches.

Why this beats prompting it yourself

Standard prompts are transient and lose detail once the context window rolls over or a new session starts. This skill automates the retrieval of specific historical facts and structural code data that manual prompting cannot reliably recall or maintain.

Use cases

• Multi-day refactoring where the agent must remember why a specific pattern was chosen across several sessions.
• Onboarding agents to large codebases by providing an indexed symbol graph and recorded architectural decisions.
• Tracking constraints to prevent the agent from suggesting solutions that were previously rejected for specific technical reasons.

Known limitations

Data paths vary between the curl installer and manual setup, requiring explicit environment variables to avoid ambiguity. Session-start context injection is limited to specific supported agents.