Engram
Nothing is forgotten.
Engram is the three-tier memory system that gives Pulse persistent knowledge across conversations. From short-term session context to permanent long-term preferences, every interaction adds to a growing understanding of who you are.
Memory operates at three temporal scales, all persisted in PostgreSQL. Session memory holds the active conversation with progressive aging — recent tool results in full, older ones compressed to save tokens. A task anchoring system re-injects the original user request after every compaction, ensuring multi-step tasks never lose their goal. Daily memory consolidates session summaries into searchable digests. Long-term memory preserves permanent knowledge: your preferences, decisions, recurring patterns, and important facts.
All three tiers are searchable through a hybrid system combining BM25 keyword matching with vector cosine similarity, backed by GIN and HNSW indexes. Engram can find memories by exact phrase, semantic meaning, or both. Learnings — patterns extracted by Forge — are stored alongside memories for compound improvement.
Memory Architecture
Session Memory
Active conversation context with progressive tool result aging. Current turn: full detail. 1-2 turns old: ~1000 chars. 3+ turns: ~200 char summaries. Originals preserved for learning.
Daily Memory
Automatic end-of-day summaries consolidate key decisions, preferences, and outcomes from your conversations.
Long-term Memory
PostgreSQL-backed permanent storage for preferences, facts, and patterns. Content-hashed for deduplication. Usage tracking per entry.
Hybrid Search
BM25 keyword matching combined with vector cosine similarity. Find memories by exact words or semantic meaning.
Contextual Injection
Relevant memories are proactively injected into every AI request. Pulse always remembers what matters.
Automatic Extraction
The Learn phase of Cortex automatically extracts and stores preferences, corrections, and patterns without manual input.
Memory Consolidation
Sessions are compacted using LLM-powered summarization. Key information is preserved while reducing token overhead.
Privacy Controls
Memory is tenant-isolated. You can view, search, export, or delete any memory at any time from the dashboard.
Memory Export
Export your entire memory as structured Markdown. Import memories from other sources. Full data portability.
How Engram remembers
Conversation happens
You chat with Pulse. Messages, tool results, and AI reasoning accumulate in session memory.
Preferences extracted
Cortex's Learn phase detects preferences, corrections, and patterns. These are stored as tagged memory entries.
Session compressed
When the session grows large, LLM summarization compresses older exchanges while preserving key information.
Daily consolidation
At the end of each day, session summaries are merged into a daily memory digest, searchable by topic and date.
Long-term learning
Repeated patterns and confirmed preferences are promoted to long-term memory. This knowledge persists indefinitely.