Detailed Introduction
DiffMem introduces a git-based differential memory approach for AI agents: current-state knowledge is stored as human-readable Markdown files while historical evolution is preserved in Git commits. Agents can load a compact “now” view for fast responses or selectively pull diffs and logs for temporal reasoning. The design emphasizes auditability, portability and token-efficient retrieval for long-horizon memory scenarios.
Main Features
- Differential view: separate surface (current files) from depth (git history) to keep active contexts lean.
- Human-readable storage: Markdown-based memory units that are easy to inspect and edit.
- Temporal queries: use git diffs and logs for time-aware retrieval and evolution analysis.
- Lightweight prototype: runs in-process with minimal dependencies (e.g., gitpython, rank-bm25) for fast experimentation.
Use Cases
- Long-lived agents: maintain auditable, reconstructable memory for agents that evolve over years.
- Research & prototyping: explore pruning, smart forgetting and temporal reasoning strategies.
- Collaborative memories: shared repos for multi-agent or multi-user memory workflows with git-based merges.
Technical Features
- Retrieval strategies: BM25 + semantic hybrid retrieval with multi-depth context assembly (core/wide/deep/temporal).
- No server required: modular library design for local in-process use.
- Extensible: combine with vector stores or external retrieval layers for enhanced semantic capabilities.
- Open-source license: MIT, suitable for research and engineering reuse.