Marrow
Sign in Create account

Agent memory for every AI.

Marrow Core keeps the context your apps and agents need as durable memory. Conversations, documents, tool calls, traces, feedback, and application records become connected facts, preferences, decisions, relationships, timing, and corrections. Provenance and time stay attached so each workflow can retrieve the relevant slice and show what it used.

Memory model.

Marrow Core operates as a high-performance knowledge engine that structures user context across three connected layers: Records (provenance-anchored documents, messages, tool calls, and activity), Graph State (connected entities, decisions, preferences, and corrections), and Context Packets (task-specific context slices returned with inspectable source lineage).

When an app or agent requires context, Marrow Core delivers the relevant documents, records, and decisions to your server or agent workflow. The hosted API accepts source URLs or files, processes them into the knowledge graph, and returns task-specific context via /v1/query—complete with source citations, confidence status, and explicit absence notices when evidence is missing.

Records

Conversations, documents, tool calls, traces, feedback, and application records, preserved with provenance and time.

Graph state

Facts, preferences, decisions, entities, relationships, timing, boundaries, and corrections connected in the knowledge graph.

Context packets

The relevant slice for the agent or app, scoped to the task and recipient.

Memory benchmarks

Marrow reaches 93.8% on LongMemEval-S, ahead of full-context and leading memory systems in published results. On LoCoMo, Marrow reaches 89.1% across multi-session recall and reasoning, shown with published results from the same systems.

LongMemEval-S

93.8%

Long-term memory accuracy

Published LongMemEval-S results Benchmark Comparison source

LongMemEval-S

Accuracy
0 100
  1. Marrow 93.8%
  2. Full context 92%
  3. Honcho 90.4%
  4. Supermemory 84.6%
  5. Zep 71.2%