Memory that thinks, not just stores
DeltaMemory goes beyond embeddings and vector search. It extracts facts, builds knowledge graphs, reasons over time, and generates insights. Your agents remember the way humans do.
Vector search finds similar text. It doesn't understand what was said.
Most memory layers store raw conversation chunks as embeddings and call it a day. Ask "where does the user work?" and you get back five paragraphs of conversation that mention jobs. Your LLM has to figure out the rest.
That works for simple cases. It falls apart when your agent needs to connect information across sessions, understand what changed over time, or answer questions that require reasoning over multiple facts.
"The user mentioned they were moving to Austin three weeks ago. Did they actually move? Where did they live before? Do they still work at the same company?"
Vector search can't answer that. Cognitive memory can.
How cognitive memory works
A single API call triggers a multi-stage cognitive pipeline. No configuration needed.
Ingest
Raw conversation text enters the pipeline. Memory is stored and embedded immediately.
Extract
Facts, concepts, events, and user profiles are identified automatically in the background.
Connect
A knowledge graph links entities and tracks how relationships evolve over time.
Recall
Hybrid search combines graph traversal, vector similarity, keyword matching, and recency scoring.
Six layers of understanding
Every feature is designed to make your agents smarter with each interaction.
Automatic Fact Extraction
When a user says "I work at Google in San Francisco and I'm allergic to peanuts," DeltaMemory extracts three distinct facts with confidence scores. No prompting required. No schema to define.
LLM-powered extraction with temporal context awareness, smart deduplication, and merge decisions across sessions.
Knowledge Graph Construction
Facts become nodes. Relationships become edges. DeltaMemory builds a semantic graph that connects concepts across conversations, enabling multi-hop reasoning that vector search alone cannot achieve.
Concept-to-concept relationships with confidence scores, bidirectional traversal, and connected concept discovery.
Reflection and Insight Generation
DeltaMemory periodically analyzes recent memories to surface patterns. It generates insights your agents can act on without being explicitly asked.
Periodic reflection over configurable windows, high-salience insight storage, and procedural pattern detection.
Three Memory Systems
Inspired by cognitive science: episodic memory stores experiences, semantic memory maintains the knowledge graph, and procedural memory captures learned patterns. Each serves a different retrieval need.
Episodic (what happened), Semantic (what is true), Procedural (how to do things).
User Profile Extraction
Structured profiles are built automatically from conversations. Topics like work, education, interests, and relationships are organized into a fast-lookup system with merge strategies.
Two-stage pipeline: Summarize then Extract. O(1) lookup by user, topic, and sub-topic.
Temporal Reasoning
DeltaMemory understands time. It distinguishes between "I worked at Google" and "I work at Google." Event timelines track when things happened, when they were mentioned, and when they changed.
82.2% temporal accuracy on LoCoMo, 23% better than the next closest competitor.
Four retrieval signals, one query
When your agent asks a question, DeltaMemory doesn't just search for similar text. It runs four retrieval strategies in parallel and fuses the results.
Vector Similarity
HNSW approximate nearest neighbor search finds semantically related memories, even when the wording is different.
Keyword Matching
BM25 full-text search catches exact terms and names that embedding models sometimes miss.
Graph Traversal
Multi-hop traversal across the knowledge graph connects facts that share entities, even across different conversations.
Recency and Salience
Recent memories and high-importance facts are boosted. Stale information naturally fades unless it keeps being relevant.
"Where does the user work now?"
3 memories about employment
2 memories mentioning companies
works_at → Stripe (confidence: 0.97)
"worked at Google" marked as past tense
Reciprocal Rank Fusion across all signals
"Stripe" — with full provenance chain
Vector search is a feature. Cognitive memory is the system.
DeltaMemory includes vector search. It also includes everything else your agents need to actually understand users.
| Capability | Plain Vector DB | DeltaMemory |
|---|---|---|
| Fact extraction from conversations | — | ✓ |
| Knowledge graph with relationships | — | ✓ |
| Temporal reasoning (past vs. present) | — | ✓ |
| Multi-hop queries across sessions | — | ✓ |
| Automatic user profile building | — | ✓ |
| Event timeline tracking | — | ✓ |
| Salience decay over time | — | ✓ |
| Similarity search | ✓ | ✓ |
| Memory consolidation and reflection | — | ✓ |
| Confidence scoring on every fact | — | ✓ |
What your agents can do with cognitive memory
Real capabilities that change how your agents interact with users.
Personalized Conversations
Your agent remembers that a user prefers concise answers, works in fintech, and mentioned a deadline next Friday. Every response is tailored without the user repeating themselves.
Proactive Suggestions
DeltaMemory surfaces insights your agent can act on. "This user has asked about pricing three times. They might be ready for a demo." Pattern detection happens automatically.
Cross-Session Continuity
A user mentions a project in January and asks about it again in March. Your agent picks up exactly where they left off, with full context on what was discussed and decided.
Temporal Awareness
"When did the user switch jobs?" "What was their opinion on X before the product update?" Your agent can reason about how things changed over time, not just what the latest state is.
Accurate Recall Under Pressure
When a user asks a question that requires connecting three different conversations from three different weeks, DeltaMemory retrieves the right facts with provenance. No hallucination from missing context.
Structured Data from Unstructured Talk
Conversations are messy. DeltaMemory turns them into structured profiles, event timelines, and fact databases. Your agent gets clean data without asking users to fill out forms.
Two API calls. Full cognitive memory.
Ingest a conversation. DeltaMemory extracts facts, builds the graph, creates profiles, and tracks events. Query it later with natural language.
The response includes extracted facts with confidence scores, identified concepts with importance weights, and memory IDs for direct retrieval.
No schema to define. No extraction prompts to write. No graph database to manage. One endpoint in, structured memory out.
// Ingest a conversation await dm.ingest('user_42', 'I just moved to Austin from NYC. ' + 'Starting a new role at Stripe next Monday.' ); // DeltaMemory extracts: // Fact: "User moved to Austin" (confidence: 0.95) // Fact: "User previously lived in NYC" (confidence: 0.92) // Fact: "User starts at Stripe" (confidence: 0.97) // Event: "New role at Stripe" (date: next Monday) // Profile: work.company = "Stripe" // Profile: demographics.location = "Austin" // Later, recall with natural language const result = await dm.recall('user_42', 'Where does the user work?' ); // Returns: "Stripe" with full provenance
Memories evolve. So does DeltaMemory.
Information isn't static. DeltaMemory handles the full lifecycle of a memory, from creation to consolidation.
Salience Decay
Memories naturally fade in importance over time, just like human memory. Frequently accessed facts stay strong. Stale information drops in priority. The decay rate is configurable per use case.
Consolidation
When similar memories accumulate, DeltaMemory clusters and summarizes them. Five separate mentions of a user's work project become one consolidated memory with the key details preserved.
Contradiction Resolution
When new information contradicts old facts, DeltaMemory detects it. "I moved to Austin" supersedes "I live in NYC." The old fact is updated, not duplicated. Your agent always has the current truth.