Skip to content

The intelligence substrate

Twenty networks, one session

To power autonomous agents, memory has to persist. ai4all integrates plug-and-play user-intelligence networks so a stateless model can behave like a continuous entity — without forcing everyone onto one schema.

20 shown

  • Vana

    Group A

    User-owned data DAOs; the wallet signs before the model reads.

    Architecture
    Local store
    Transport
    stdio

  • Silmari

    Group A

    Active forgetting as a first-class primitive, executed on device.

    Architecture
    Local store
    Transport
    stdio

  • Personal AI

    Group A

    A personal language model trained on one person's own corpus.

    Architecture
    Hybrid
    Transport
    stdio

  • Rewind

    Group A

    Continuous local capture of everything seen, said, and heard.

    Architecture
    Capture stream
    Transport
    stdio

  • Tenure

    Group A

    Durable local memory tenure without a cloud round-trip.

    Architecture
    Local store
    Transport
    stdio

  • MemTrust

    Group A

    Attested memory: every recall carries a verifiable provenance chain.

    Architecture
    Structured graph
    Transport
    stdio

  • Dina

    Group A

    Credentialing and proof-of-competence issuance for agents and humans.

    Architecture
    Structured graph
    Transport
    stdio

  • OpenMemory

    Group A

    Open-source, self-hosted memory layer shared across local agents.

    Architecture
    Flat vector
    Transport
    stdio

  • Mem0

    Group B

    Fast flat-vector memory with ultra-low read latency.

    Architecture
    Flat vector
    Transport
    http-stream

  • Letta

    Group B

    OS-style virtual memory paging for long-lived agents.

    Architecture
    Hybrid
    Transport
    http-stream

  • Wispr Flow

    Group B

    Voice-first capture that turns speech into structured context.

    Architecture
    Capture stream
    Transport
    http-stream

  • Zep

    Group B

    Bi-temporal knowledge graph; facts carry valid-time and ingest-time.

    Architecture
    Structured graph
    Transport
    http-stream

  • Cognee

    Group B

    ECL pipelines that cook raw documents into a queryable memory graph.

    Architecture
    Structured graph
    Transport
    http-stream

  • Hypabase

    Group B

    Hyperedge indexing for dense, multi-role facts.

    Architecture
    Structured graph
    Transport
    http-stream

  • Memanto

    Group B

    Flat-vector recall tuned for rapid, zero-ceremony ingestion.

    Architecture
    Flat vector
    Transport
    http-stream

  • A-MEM

    Group B

    Agentic memory that reorganises its own note network over time.

    Architecture
    Structured graph
    Transport
    http-stream

  • Hindsight

    Group B

    Retrospective reasoning over what the agent already did.

    Architecture
    Hybrid
    Transport
    http-stream

  • Theoria

    Group B

    Hypothesis-shaped memory for multi-agent research workflows.

    Architecture
    Structured graph
    Transport
    http-stream

  • Granola

    Group B

    Meeting context that survives the meeting.

    Architecture
    Capture stream
    Transport
    http-stream

  • Touchy

    Group B

    Relationship memory: who, when, and what was promised.

    Architecture
    Flat vector
    Transport
    http-stream

Why both sides

The trade-off is real, so the hub absorbs it

Flat-vector and structured-graph memory fail differently. Neither is wrong; they suit different workloads. Routing between them is the job.

Flat vector

e.g. Mem0, Memanto, OpenMemory, Touchy

Latency
Very low. Ingestion is cheap and near-instant.
Drawback
Token bloat. Similarity search returns approximately-relevant spans, and over a long history that compounds.
Long-conversation precision
Degrades. With no relational structure, the right fact and a nearby wrong one look alike.

Structured graph

e.g. Zep, Hypabase, Cognee, A-MEM

Latency
Higher on the write path — extraction and linking cost something.
Advantage
Dense, multi-role facts in hyperedges with bi-temporal indexing: what was true, and when we learned it.
Long-conversation precision
Higher. Structure reduces contextual fragmentation and the hallucination risk that follows it.

Neutrality

The membrane

Group A · SovereignLOCAL-FIRST · CRYPTOGRAPHIC · STDIOGroup B · Non-sovereignCLOUD-COORDINATED · HTTP STREAMVanaSilmariPersonal AIRewindTenureMemTrustDinaOpenMemoryMem0LettaWispr FlowZepCogneeHypabaseMemantoA-MEMHindsightTheoriaGranolaTouchyai4all core routing

The memory sector has bifurcated. A hub that picks a winner is just another walled garden — so the routing seam addresses both sides through one session.

Protocol pillars

What the host speaks

Seven open specifications and their stewards. Listed to describe the interfaces ai4all implements — not as claims of partnership or endorsement.