Memory validation
for AI agents.
RAVEN is a seven-engine memory validation pipeline. It doesn't compete with retrieval systems — it runs after them. Every memory passes through entity normalization, causal chain construction, temporal decay, contradiction detection, importance ranking, and a confidence gate before the agent sees it.
Retrieval and validation are different layers.
Systems like MemPalace store verbatim and find later via semantic search. They return chunks — including contradictions, stale facts, and importance-inverted noise.
The retrieval layer trusts the agent to sort the results.
RAVEN runs after retrieval. Every returned memory passes through the seven-engine pipeline. Contradictions are flagged. Low-confidence results are refused. The agent only sees validated memories.
The validation layer sorts before the agent sees.
Open source.
MIT licensed. Repository is public.
Deterministic heuristics (v1).
Engines are heuristic. No LLM calls at runtime. Fully reproducible.
v2 roadmap.
LLM-inference upgrades to NOVA (causal) and PULSAR (contradiction) planned.