Address a memory. Get a grounded answer.

Meo is a memory agentic layer. A query is a current through short-term cores, the knowledge graph, and the context window — then the answer lands.

Chat is the probe

Ask in the workspace. Meo rewrites the query, walks STM and the graph, then answers with a visible trace — not a black box.

workspace · live recallsynthetic demo
What do I remember about my dog?

Recalled from short-term memory and the graph: your dog is Miso, a rescue who hates vacuum cleaners and waits by the door when you say “walk.”

ReceivedEditingGreppingReadingGrounding
Ask a follow-up…

The current is the product

Seven stages, one path. In Workflow you watch the job move: query in, memory written back.

  1. 000ReceivedQuery accepted
    001ThinkingGuardrails + intent
    002EditingQuery rewrite
    003GreppingShort-term memory
    004ReadingKnowledge graph
    005GroundingContext window
    006DoneRecalled
workflow · current#a12f9c
ReceivedQuery accepted
ThinkingGuardrails + intent
EditingQuery rewrite
GreppingShort-term memory
ReadingKnowledge graph
GroundingContext window
DoneRecalled

Short-term ring

open memory →

Redis keeps the last exchanges as a ring. Newest cores sit at the front of the tape.

  1. user#4

    What do I remember about my dog?

  2. agent#3

    Miso — rescue, hates vacuums, waits for walks.

  3. user#2

    Does my friend love birds?

  4. agent#1

    Yes — Priya loves birds; noted on the graph.

Knowledge graph

open knowledge →

Long-term memory is Neo4j: names, types, and edges written after a recall — not a chat log.

  • PriyaLOVESbirds
  • youHASMiso
  • MisoTYPEdog

Context window

open context →

Before Meo answers, it reads this sliding window — markdown slots that ground the next turn.

Window utilisation3 / 8
slot 3grounding

Miso

  • Rescue dog; dislikes vacuum cleaners
  • Responds to “walk”

Put a current through your memory.

Open the workspace, ask something Meo should already know, and watch STM, the graph, and context take the load.

Open workspace