Skip to content
BrainAPIBenchmarks
Star545Open BrainAPI

01 / 06 · Memory layer

Graph-backed memory that holds up on long-context recall

BrainAPI turns unstructured text into an event-centric knowledge graph. You ask in plain language and get answers along a walkable path — not a similarity guess. On published LoCoMo and BEAM1M scores, it leads Mem0, GetZep, and Letta.

Open BrainAPIView GitHub
LoCoMo · BrainAPI
95.39%
BEAM1M · BrainAPI
78.97%

Every action is an event hub: who acted, on what, where, and when.

02 / 06 · LoCoMo

Long-conversation memory, ranked

Long-conversation memory. BrainAPI leads published scores against Mem0, GetZep, and Letta.

View results and reports

  1. 01BrainAPI

    95.39%

    BrainAPI scored 95.39% on LoCoMo.

  2. 02Mem0

    92.5%

    Mem0 scored 92.5% on LoCoMo.

  3. 03GetZep

    80.32%

    GetZep scored 80.32% on LoCoMo.

  4. 04Letta

    74%

    Letta scored 74% on LoCoMo.
LoCoMo scores for BrainAPI, Mem0, GetZep, and Letta.
RankSystemScoreStatus
1BrainAPI95.39%Highest published
2Mem092.5%Published
3GetZep80.32%Published
4Letta74%Published

03 / 06 · BEAM1M

Million-token horizon recall

Million-token horizon recall. BrainAPI and Mem0 have published scores; GetZep and Letta have not.

View results and reports

  1. 01BrainAPI

    78.97%

    BrainAPI scored 78.97% on BEAM1M.

  2. 02Mem0

    64.1%

    Mem0 scored 64.1% on BEAM1M.

  3. 03GetZep

    Not published

    GetZep has no published BEAM1M score.

  4. 04Letta

    Not published

    Letta has no published BEAM1M score.
BEAM1M scores for BrainAPI, Mem0, GetZep, and Letta.
RankSystemScoreStatus
1BrainAPI78.97%Highest published
2Mem064.1%Published
3GetZepNot publishedNot published
4LettaNot publishedNot published

04 / 06 · Event graph

Facts stay as events, not flattened edges

BrainAPI is event-centric. Instead of storing only that A is related to B, it captures who did what, to whom, when, and in what context. That structure is what makes multi-step questions answerable.

  • Actor, target, and context hang off a central event node, so provenance survives every ingest.
  • The graph is append-only. New actions are added as their own events rather than overwriting what was already known.
  • You can reconstruct how a fact evolved because events are never merged away.

05 / 06 · Retrieval

Answers come with the path that produced them

Retrieve context in plain language. The response includes the graph triples used to derive it, so you can inspect the walk instead of trusting a nearest neighbour.

result = client.retrieveContext("Who organized AI events in London in Q1 2024?")
# text_context + triples + historical_context

06 / 06 · Agent swarm

A swarm reads, notes, and draws the graph

Each ingest is handled by specialized agents with one narrow job, which keeps the pipeline auditable from raw text to stored triples.

  1. 01
    Scout

    Extracts people, things, events, and quantities from the text.

    Extract
  2. 02
    Observations

    Writes notes about the facts using previously known context.

    Note
  3. 03
    Architect

    Builds event-centric relationships around those entities.

    Connect
  4. 04
    Janitor

    Verifies correctness, catches duplicates, and loops until the graph is clean.

    Verify

Next step

Run the memory layer on your own text

Open the product, or clone the source and start from the TUI. Persistent structured memory with a traceable answer path is the point.

Open BrainAPIGitHub repository

Reach for BrainAPI when you need structured knowledge from messy sources, long-term explainable memory, or multi-hop queries with a traceable answer.