Graph + vector engine with episodic, semantic, and procedural memory. Agents that remember, learn, and evolve.
Perspective separates memory the way brains do. Each type decays, consolidates, and retrieves differently.
Specific events with timestamps. "Yesterday I discussed the trading strategy." Decays fast unless referenced.
Extracted facts and knowledge. "Bodmash prefers concise responses." Slow decay, strengthened by reinforcement.
Skills and action patterns. "When reviewing PRs, check for force pushes first." Never decays.
Vector similarity, full-text search, and graph traversal fused with Reciprocal Rank Fusion. No single-modality blind spots.
score = recency(x) × importance(x) × relevance(x, query)
// recency — exponential decay by memory type half-life
// importance — access_count + user weight + extraction confidence
// relevance — RRF(vector cosine, BM25, entity graph, temporal proximity)Zero-copy, lock-free. Embedded Qdrant Edge means zero Docker dependency.
Full isolation. No shared collections, no logical filtering bugs.
Smart batching. Importance gate filters noise. Generic OpenAI-compatible API.
Memories fade unless accessed. Principled forgetting, not arbitrary TTL.
Dedup, community detection, episodic-to-semantic promotion. Runs like sleep.
First-class plugin. Drop-in replacement for any Hermes memory provider.
Fast, typed, streaming. Embed as a library or run as a service.
PyO3 bindings. `pip install perspective-core`. Rust engine, Python API.
LangChain, CrewAI, AutoGen, custom. Any agent, any framework.
Self-hosted is free. Cloud is pay per use.
MIT license. Your infra.
Managed. Scale without ops.
Multi-agent, SLA, private
Join the waitlist for Perspective Cloud. Early access starts Q3 2026.