Comparison
AIDE Memory is built to solve the core problems in existing memory tools.
Feature Comparison
| Feature | AIDE Memory | claude-mem | engram |
|---|---|---|---|
| Adoption mechanism | 4 hooks (100% automatic) | System prompt injection | None (0% voluntary) |
| Token cost per recall | ~20 tokens (nudge) | ~2,000 tokens (full dump) | Variable |
| Path scoping | Glob inheritance from day one | Folder Context Files (recent) | None |
| Storage | JSON files + SQLite cache | ChromaDB + SQLite | SQLite |
| Infrastructure | npx, nothing else | Docker + ChromaDB + HTTP server | Go binary |
| Memory reliability | 544 tests, 0 controllable failures | 72% summary failure rate | Empty/ghost observations |
| File pollution | One directory: .aide/ | Creates files in every directory | Clean |
| Security | Local-only, no network | Unauthenticated HTTP API on port 37777 | Windows Defender flags binary |
| Git sync | Native (memories are files) | No | No |
| Compaction protection | PreCompact hook saves decisions | No | No |
| Structured layers | 4 layers (preferences, technical, area_context, guidelines) | Flat | Flat |
| License | FSL (auto-converts Apache 2.0) | AGPL-3.0 | MIT |
| Price | Free | Free | Free |
Why AIDE Memory is Different
Adoption, not availability. When memory tools are simply available as optional MCP tools, agents use them 0 out of 10 times. AIDE Memory uses hooks to capture memories at the moment they happen — no user intervention required.
Token efficiency. Dumping all memories into the system prompt on every interaction costs ~2,000 tokens. AIDE Memory sends a 20-token nudge and lets the agent decide relevance. Over a full day of work, this saves tens of thousands of tokens.
Path scoping from day one. A memory about test utilities shouldn’t appear when you’re working in database code. AIDE Memory uses glob inheritance to surface memories only where they matter.
Git-native sync. No Docker. No vector database service. No API keys. Memories are files. Files are committed. Git syncs them. Team members automatically get shared memories via git pull.
Compaction protection. Context compaction is documented as a source of lost decisions across 350+ GitHub issues in existing tools. AIDE Memory has a PreCompact hook that extracts planning decisions before they’re lost.
Structured layers. Preferences, technical facts, area context, and team guidelines are stored separately, so recall can filter by layer and avoid noise.