Skip to main content
Engra
Memory

Agent Memory

Long-term memory production agents remember—with predictable token spend

Organize enterprise knowledge as governed memory units with clear team, domain, and topic boundaries. Layered wake-up loads core context at session start and expands on user questions—less repetition, fewer off-topic answers, and no linear token bill growth every chat.

Capabilities

  • ✓Multi-team governance: Library / Scope / Topic memory boundaries
  • ✓Layered wake-up: core context at start, details on demand—predictable token cost
  • ✓Traceable links: path backlinks for audit and knowledge evolution
  • ✓Published benchmark: ENGRA-KB-v2 Recall@10 75.3% (+20pp vs chunk RAG)
  • ✓Production access: Memory MCP + REST API, quotas and SLA

Memory hierarchy

Library → Scope → Topic → Atom answers: which workspace, which domain, which subject, and the concrete retrievable unit.

  1. 1

    Library

    Isolated memory workspace (tenant boundary): separate storage, vectors, and indexes. Each team can create multiple libraries per project or use case.

    Examples: support knowledge base, engineering runbooks, onboarding library

  2. 2

    Scope

    Broad thematic bucket for work domains—not a single memory title. The system can suggest a scope on write when left empty.

    Examples: engineering, finance, operations

  3. 3

    Topic

    Narrow subject within a scope (billing, auth, releases). Classified by meaning, not calendar buckets by default; encode time periods in names if needed.

    Examples: billing, api-auth, onboarding

  4. 4

    Atom

    Smallest retrievable unit: structured body text, optional source attachment, and a semantic vector. Cross-memory citations carry Scope / Topic / title paths.

    Examples: a billing rule, an incident postmortem summary

Memory stack (wake-up & retrieval)

Works with the four layers: what to load when a session starts vs. what to search during conversation.

  • L0L0 Identity — stable self-description
  • L1L1 Core narrative — top salient atoms
  • L2L2 Recall on demand — filter by Library / Scope / Topic
  • L3L3 Deep search — semantic search across the library

Product model is Library / Scope / Topic / Atom. Public API JSON may still expose persona bound to the library (library-{id}); see API docs for field details.

Backlinks graph

2D shows linked atoms on one plane; 3D reveals compressed topic clusters behind—re-summarized from correlated atoms.

[[/finance/billing/invoice-rules: Due on the 25th each month]]

2D

Interactive demo with sample data. Your team's live graph appears in the console after atoms are linked.

Benchmarks & published results

On ENGRA-KB-v2 (550 docs / 400 queries), Structured atom retrieval beats BM25 and traditional chunk-RAG — full tables and downloadable JSON.

76.0%

Recall@10

+24.0 pp vs BM25

0.301

MRR

+17% vs BM25

550/400

Docs / queries

ENGRA-KB-v2 paper-scale suite

Synthetic enterprise corpus · same embedder · retrieval stage only

View test resultsFull methodology →

API prefix

/api/v1/memory
API docsConsole