Write-optimized control plane that agents inhabit as a runtime contract — durable handoff over amnesia
Most agent memory systems optimize the read path: Mem0 for knowledge graphs, Zep for temporal reasoning, Hindsight for multi-strategy retrieval. ContextLattice uniquely optimizes the write path.
Context freshness under load: In long-horizon agents, write latency cascades → queue congestion → stale retrieval indexes → poor context → unreliable decisions. ContextLattice explicitly models the efficacy chain:
Write speed ↑ → queue pressure ↓ → sink freshness ↑ → retrieval freshness ↑ → recall quality ↑ → task completion reliability ↑
┌─────────────────────────────────────────────┐
│ Single Orchestrator Entry Point │
│ - Durable raw write (Mongo ledger) │
│ - Async fanout to specialized sinks │
│ - Admission control + queue monitoring │
└──────┬──────────────────────────────────────┘
│
├─→ Fast Sinks: Topic rollups, Qdrant, pgvector
├─→ Deep Sinks: MindsDB, Letta, memory-bank
└─→ Durable Outbox: Coalescing, retries, backpressure
Fast Lane (p50: 450ms-2s) ├─ Topic rollups (high-signal summaries) ├─ Qdrant (semantic vectors) └─ PostgreSQL-pgvector (dense embeddings) ↓ Fusion + Learning Reranking ├─ Multi-source result merging ├─ Feedback-driven ranking improvement └─ Code-context enrichment (symbol overlap, file-path proximity) ↓ Deep Continuation (async, non-blocking) ├─ MindsDB (ML/analytics queries) ├─ Mongo raw (full ledger search) ├─ Letta (agent memory integration) └─ memory-bank (specialized retrieval)
Fail-open degradation: Returns fast-lane results even if deep lanes timeout. Retrieval never blocks on slow backends.
v3.4.0 (2026-06-05) crosses CL from a callable memory service to an inhabitable agent runtime. The website headline shifted from write-optimization / freshness framing to amnesia prevention: "Stop giving your agents amnesia and calling it workflow."
contextlattice_agent_adapter lifecycle — bootstrap, context-pack, checkpoint, handoff, event, complete — that any agent runtime implements once./v1/agents/sessions with /telemetry/agents/runtime, plus objective_runtime_state.v1 threaded across preflight / context-pack / Dream Mode.Gateway: Go service (41 files, 212KB) — HTTP ingress, write admission, retrieval coordination
Core Engine: Rust crates — context_engine (memory graph primitives), context_retrieval (ranking/indexing), context_codec (serialization)
Vector Stores: Qdrant, Weaviate, LanceDB backends
Relational: PostgreSQL-pgvector, MongoDB, MindsDB
LLM Routing: Ollama (local models), Letta integration
Protocol: MCP-native (Model Context Protocol, 100 msgs/sec writes)
| System | Primary Focus | Write Path | Read Path |
|---|---|---|---|
| ContextLattice | Context freshness + efficacy | Durable orchestrator + fanout | Staged multi-lane + learning rerank |
| Letta | Stateful agent memory blocks | Agent tools (editable) | In-context injection |
| Mem0 | Structured knowledge graphs | Vector + entity extraction | Graph + vector search |
| Zep | Temporal knowledge graphs | Entity extraction | Temporal graph traversal |
| Hindsight | Multi-strategy retrieval | Standard RAG | 4 parallel strategies (semantic, BM25, graph, temporal) |
ContextLattice's niche: Long-horizon agent reliability where write latency matters. Less useful if you need temporal reasoning (Zep) or knowledge graphs (Mem0/Cognee).
Isolation mechanism: None / Prompt-level on the spectrum. Local-first, API-key auth on protected endpoints, HTTP+MCP transport only. No container, micro-VM, or network-level isolation.
Multi-tenancy trade-offs: Single-node local-first by design — there is no cross-tenant firewall because there is no multi-node tenancy model. The boundary is the box it runs in.
Recommended architecture: Single-tenant on-premises for air-gapped deployments; any multi-tenant SaaS use must layer RBAC and process isolation on top.
Active deployments: Payment processing features (Stripe webhooks, PayPal verification) indicate real paying customers.
Release velocity: Very high — 30 GitHub releases, v3.3.x through v3.4.23 in roughly Feb–June 2026; repo pushed 2026-06-14, not archived.
Community: 113 stars, 7 forks (small but focused), registered on Glama MCP server platform.
License: Business Source License 1.1 (non-open, commons clause).
Version boundary: v3.4.0 is the stable public agent-runtime-contract baseline; v4 remains a private tuning/experiment lane gated on benchmark/recall/soak.
Performance baseline concerns: Public perf-baseline.md shows degradation on complex workloads (multi-agent: p50=21s with 33% timeout rate). No external benchmarks comparing retrieval efficacy against competitors.
Choose ContextLattice if:
Choose alternatives if: