01 · graph
Knowledge graph
Store people, companies, projects, preferences, meetings, and ownership lines as linked entities instead of loose transcript fragments.
Built on Helix-DB · Graph + Vector
Fast Memory
for AI Agents
Supersalience turns messages, events, and docs into a typed knowledge graph your agent can recall before the next model call. Use one API for implicit learning, graph-vector search, and production-safe memory policy.
100%
Pilot accuracy
10ms
Mean search
2393
Context tokens
01 // What it does
Stop rebuilding context inside every prompt.
02 · search
Semantic search
Search text, vectors, contacts, and seed entities together, then expand the graph just far enough to recover hidden connections.
03 · memory
Durable memory
Keep per-user facts with category, source, confidence, stability, importance, durability, and policy reasons attached.
04 · sessions
Session continuity
Carry important entities across turns and sessions without adding app-side memory caches to every product surface.
05 · ttl
Policy-gated learning
Gateway-native extraction rejects probe, setup, request-meta, and weak inferred facts before they become permanent state.
06 · scale
Built on Helix-DB
Run on Rust infrastructure with graph storage and ANN vector search powered by Helix-DB.
02 // Why it exists
Agents fail when memory is raw chat history.
problem
Chat logs bury facts
A user can mention a buyer, launch freeze, or preference once. The agent still needs that fact weeks later without replaying the whole conversation.
risk
Vectors miss ownership
Similarity search can find nearby text, but it does not preserve who owns compliance, which account is at risk, or which preference applies to which customer.
control
Naive extraction pollutes state
Supersalience scores, rejects, dedupes, and reranks facts before they reach memory, so temporary tasks do not become long-lived beliefs.
03 // Quickstart
Recall before the model. Learn after the answer.
POST /memory/context
# Hot-path memory context before a chat turn curl -X POST https://supersalience.com/memory/context \ -H "authorization: Bearer $SUPERSALIENCE_TOKEN" \ -H "content-type: application/json" \ -d '{ "user_id": "demo-user", "session_id": "demo-chat", "query": "what should I know before the Acme renewal call?", "include_knowledge": false, "limit": 18, "max_scan": 30, "max_entities": 12 }'
# Queue implicit extraction after the assistant answers curl -X POST https://supersalience.com/memory/implicit-learn \ -H "authorization: Bearer $SUPERSALIENCE_TOKEN" \ -H "content-type: application/json" \ -d '{ "user_id": "demo-user", "session_id": "demo-chat", "source_message_id": "msg_123", "user_text": "Mira owns compliance for Acme and hates Friday launches.", "assistant_text": "Noted. I will avoid Friday launch plans for Acme.", "model": "openai/gpt-4.1-mini" }'
# Show the memories currently accepted for a user curl -X POST https://supersalience.com/memory/list \ -H "authorization: Bearer $SUPERSALIENCE_TOKEN" \ -H "content-type: application/json" \ -d '{ "user_id": "demo-user", "limit": 50, "offset": 0 }'
# Persist typed graph observations directly curl -X POST https://supersalience.com/sync \ -H "authorization: Bearer $SUPERSALIENCE_TOKEN" \ -H "content-type: application/json" \ -d '{ "user_id": "demo-user", "entities": [ { "name": "Mira Patel", "entity_type": "person", "observations": ["Owns compliance for Acme."] }, { "name": "Acme", "entity_type": "organization", "observations": ["Renewal risk depends on compliance review."] } ], "relations": [ { "from_name": "Mira Patel", "to_name": "Acme", "relation_type": "owns_compliance" } ], "refresh_search_state": true }'
# Search graph + vector + text with session carry-over curl -X POST https://supersalience.com/search \ -H "authorization: Bearer $SUPERSALIENCE_TOKEN" \ -H "content-type: application/json" \ -d '{ "user_id": "demo-user", "session_id": "demo-chat", "queries": [ { "query": "Acme renewal compliance owner launch preference", "weight": 1.0 }, { "query": "Mira Acme Friday launch risk", "weight": 0.7 } ], "limit": 10, "depth": 1, "max_entities": 50 }'
04 // MemoryBench
Proof you can inspect, not a demo claim.
score
100% / 10ms / 2393tok
MemScore from the 5-question LongMemEval-S pilot. The run is preserved in the benchmark docs and reproducible from the CLI.
accuracy
5 / 5
The pilot answered every sampled LongMemEval-S question correctly. This is a pilot gate, not a full 500-question leaderboard run.
latency
10ms search
Mean search latency across the pilot. Mean ingest was 26.1s per question because extraction and bulk sync happen before recall.
ingest
Bulk, then refresh
Sessions become dated source chunks. Atomic facts are extracted, written through `/sync` with deferred refresh, then reindexed once with `/admin/recompute-search-state`.
retrieve
Wide fanout
Search asks for top-50 internal candidates, expands deterministic query chunks, filters source chunks, then reranks compact facts down to the top 10.
diagnose
Question traces
Every MemoryBench question can emit expanded queries, raw top-50, reranked top-10, expected-session hit flags, and final answer evidence.
05 // FAQ
The objections that matter in production.
01 · scope
Is this just vector search?
No. Supersalience combines graph entities, relations, observations, lexical matching, vector search, session seeds, and deterministic reranking.
02 · latency
Does learning block chat?
No. Recall stays on the hot path. Implicit learning is queued after the assistant response and acknowledged quickly by the gateway.
03 · control
How do you avoid bad memories?
Every candidate gets policy-scored before write. Generic tasks, probe facts, request meta, setup chatter, and weak inferences are rejected before storage.
06 // Powered by
Storage layer
helix-db.com
The infinitely scalable graph-vector database, built in Rust. Supersalience is a typed knowledge and memory layer on top of helix-db, tuned for AI agent runtimes.