Stop wasting tokens and time on non-deterministic vector searches for literal stack traces. ogm-lightweight delivers 100% explainable recall using FTS5 BM25, exact term triggers, and temporal graph traversal in a single Go binary.
ogm-lightweight structures persistent agent context into six canonical abstractions for 100% explainability.
Immutable raw observation capturing error logs, terminal outputs, code diffs, and verification outputs.
Durable conclusion backed by evidence episodes, containing confidence scores, status, and version history.
Named codebase item (symbol, repository, file, package) with temporal version tracking.
Typed connection linking entities together (e.g., checkout.go → SQLITE_BUSY → Bounded_Retry).
Scoped key-value knowledge attached directly to memories for instant metadata filtering.
Registry of allowed entity types and relation predicates enforced across agent sessions.
AI agents interact with ogm-lightweight via six standardized Model Context Protocol (MCP) tools.
Searches memories using FTS5 BM25, exact term triggers, and graph traversal. Returns ranked exact matches in <1.2ms.
Ingests raw observations (terminal command outputs, compiler errors, code changes) as immutable evidence episodes.
Saves durable memory conclusions linked to supporting episode IDs with assigned confidence scores.
Allows agents to correct, invalidate, or adjust stale conclusions when new workspace evidence emerges.
Soft archives obsolete memories by default. Explicit hard-delete requires project owner token credentials.
Returns complete stored provenance, explain scores, and confidence fields without revealing configured secrets.
Watch how ogm-lightweight ingests raw observations, redacts sensitive API credentials, indexes structured memory episodes into SQLite STRICT, and performs multi-signal recall.
Agent captures raw code edits, stack traces, and command logs via memory_observe.
Regex & AST sanitizer scrubs OpenAI API keys, JWT tokens, and DSN passwords automatically.
SECURITY FIRSTAtomic transaction indexes text with FTS5 BM25 and updates temporal graph relations.
ZERO-CGO STORAGECombines BM25 weights, exact trigger matches, and graph depth-2 traversal for exact solutions.
1.1ms RECALLClick any node below to inspect relationships between code repositories, files, error episodes, conclusions, and traits.
Install agent-side MCP connectors and skill policies in one command using the native ogm-lm CLI utility.
Automatically configures environment variables in ~/.config/ogm-lightweight/env with 0600 permissions.
# Short aliases for quick harness setup ogm-lm --h opencode --apply ogm-lm --h claude-code --apply
Verify project API keys, selected harness target, skill paths, and exposed MCP tool lists without modifying state.
# Safe preview & diagnostic modes ogm-lm harness doctor --harness opencode ogm-lm harness print --harness claude-code
Adjust the multi-signal sliders to observe how ogm-lightweight computes composite memory relevance scores deterministically.
Calculating...Test ogm-lightweight's real-time regex & AST sanitizer. Type or select a preset to see secrets scrubbed before database persistence.
Compare ogm-lightweight against Pinecone / Vector DBs and Naive RAG setups for AI coding agent persistent memory.
| Metric Feature | ogm-lightweight | Vector DB (Pinecone / Qdrant) | Naive File RAG |
|---|---|---|---|
| Embedding API Cost | $0.00 (Zero Cost) | $500+/mo OpenAI | $0 - $50/mo |
| Stack Trace Recall | 100% Deterministic | ~65% (Fails line numbers) | ~50% (Token truncations) |
| Recall Latency | < 1.2 ms | 120 - 250 ms | 45 - 90 ms |
| Memory Footprint | 28 - 35 MB RAM | 4 GB+ RAM | 500 MB RAM |
| Architecture | 1 Binary, 1 SQLite File | Multi-container Cloud | Python + FAISS |
Select your AI coding client to copy the auto-generated MCP configuration.
This static website and REST/MCP server endpoints are reverse-proxied using Caddy on ogm-lightweight.svclabs.cloud.
ogm-lightweight.svclabs.cloud, http://ogm-lightweight.svclabs.cloud { root * /var/www/ogm-lightweight file_server encode gzip }
[Unit] Description=OGM Lightweight Memory Engine After=network.target [Service] ExecStart=/usr/local/bin/ogm-lm serve --port 8080 Restart=always User=ogm