context-m
Deterministic agent memory system that stores facts without LLM calls, supports Mem0 compatibility, and guarantees byte-exact reproducibility across runs.
Ficha actualizada 23 sept 2026, 7:31 · upstream 23 sept 2026, 7:19 · release v0.6.0
Qué es context-m
Descripción editorial construida a partir del repositorio, su documentación y sus señales de mantenimiento.
CortexM is a deterministic agent memory framework designed to store and retrieve facts with zero LLM dependency. It uses a hash-based chaining mechanism where every retrieved fact carries a BLAKE3 hash back to its source text, ensuring complete determinism. The system supports Mem0-compatible drop-in replacement, operates entirely offline without API keys or subscriptions, and guarantees the same result on every invocation. Each stored fact is indexed by user ID, timestamp, and content, with retrieval returning only verified facts that can be derived from the provided context.
- Store and retrieve factual knowledge for agent memory without LLM calls
- Replace expensive external knowledge sources (e.g., Wikipedia) with local deterministic storage
- Provide consistent fact retrieval across multiple sessions or agents
- Enable offline operation by storing facts in a single file without API keys
- Support temporal reasoning through calendar-aware chunking and date resolution
- Serve as a drop-in replacement for Mem0-compatible memory systems
- Deterministic: every retrieval produces the same byte-exact result across runs with identical inputs
- Zero cost: no API keys, no subscriptions, $0 monthly usage on any hardware
- Mem0 compatible: can be dropped-in as a drop-in replacement for existing Mem0 implementations
- BLAKE3 chaining: every retrieved fact carries its source hash chain back to the original text
- Temporal reasoning support: handles relative time expressions and calendar-aware date resolution
- Subtask-level verification: full 500-question benchmark achieves 1.000 (500/500) with provenance-stamped results
- Single `.db` file storage means all facts are stored in one location, which may become large for very long-running agents
- No active learning: the system does not automatically update its knowledge base without explicit user intervention or re-ingest
- Deterministic but static: while results are reproducible, the agent cannot learn from new information unless explicitly updated
- Limited to single-user per session: facts are scoped to a specific `user_id` and must be retrieved by that same identifier
- No multi-agent coordination: each Memory instance operates independently without inter-process communication
Perfil de infraestructura
Estimación inicial derivada del repositorio y de los servicios detectados.
Despliegue detectado
Qué señales reales de instalación hemos encontrado.
Install via pip with `pip install cortexm`. Run in production mode with `cortexm --prod`, which disables debug output. For Docker, use a custom image or run the binary directly. The framework supports both Python and Rust backends. Results are stored as JSON files under `benchmarks/results/` directory. To verify determinism, run the same query twice and compare SHA-256 hashes of retrieved results.
Stack y dependencias
Componentes inferidos a partir de manifests, configuración y documentación.
Puede sustituir a
Relaciones detectadas por el análisis del repositorio y su categoría.
Experiencia real de la comunidad
Ayuda a contrastar la estimación automática con instalaciones reales.
Todavía no hay datos comunitarios para este proyecto.