SynapCores Memory

Long-term memory for the agent, stored in a database you run yourself. Memories are embedded on write and retrieved by meaning, not keywords — "what coffee does he like" finds "prefers dark roast" without sharing a single word.

Nothing leaves the machine. Embeddings are computed inside the engine; there is no external API key and no third-party service in the path.

When to use this

• The user says "remember that…", "don't forget…", "for future reference…"
• The user asks "what do you know about…", "what did I tell you about…"
• The user corrects or retracts something you stored
• You need context from an earlier session

Setup

1. Run a SynapCores gateway (once):

docker run -d --name synapcores -p 8080:8080 \
  -v synapcores-data:/var/lib/synapcores \
  -e SYNAPCORES_TELEMETRY=off \
  ghcr.io/synapcores/community:latest

The first boot prints an admin login and an API key:

docker logs synapcores 2>&1 | grep -A2 'ADMIN LOGIN'

2. Export the key:

export SYNAPCORES_API_KEY="aidb_…"
export SYNAPCORES_URL="http://localhost:8080"

3. Check it answers:

curl -s "$SYNAPCORES_URL/health"
# {"status":"ok","timestamp":"…"}

Usage

Every operation is one SQL statement over HTTP. Define this helper once per shell session:

sc() {
  curl -s -m 120 -X POST "$SYNAPCORES_URL/v1/query/execute" \
    -H 'Content-Type: application/json' \
    -H "Authorization: Bearer $SYNAPCORES_API_KEY" \
    -d "$(jq -nc --arg sql "$1" '{sql:$sql}')"
}

Using jq -nc to build the body matters: memory text routinely contains quotes and apostrophes, and hand-built JSON breaks on them.

Remember something

sc "SELECT MEMORY_STORE('default', 'Luis prefers dark roast and takes calls after 10am') AS id" \
  | jq -r '.data.rows[0][0]'
# mem_1kzv0fkzy_AK7TUf

Note the SQL string quoting: single quotes are strings, double quotes are identifiers. A double-quoted value is read as a column name and will error. Escape an apostrophe by doubling it — 'Luis''s laptop'.

Optional metadata for later filtering:

sc "SELECT MEMORY_STORE('default', 'Renewal call scheduled for March',
      json_object('category','decision')) AS id" | jq -r '.data.rows[0][0]'

Recall

sc "SELECT content, similarity FROM MEMORY_RECALL('default', 'what coffee does he like', 5)" \
  | jq -r '.data.rows[] | "\(.[1] | tonumber | .*100 | round)%  \(.[0])"'
# 39%  Luis prefers dark roast and takes calls after 10am

That is the whole point: the stored sentence never says "coffee", and it is still the top hit. Keyword search returns nothing here.

MEMORY_RECALL(namespace, query [, top_k]) is table-valued — it goes in FROM, not SELECT. It returns (id, content, similarity, metadata, created_at), ordered by similarity, top_k defaulting to 10 and capped at 100. An unknown namespace returns zero rows rather than an error.

Read the similarity before you trust a hit. These are cosine scores, and a genuine match on this embedding model often sits around 0.5 rather than near 1.0. A low score across every row means nothing relevant is stored — say so instead of presenting the closest row as fact.

Forget

ID=$(sc "SELECT id FROM MEMORY_RECALL('default', 'renewal call', 1)" | jq -r '.data.rows[0][0]')
sc "SELECT MEMORY_FORGET('default', '$ID') AS deleted" | jq -r '.data.rows[0][0]'
# true

Returns true if a row was deleted, false if the id did not exist. This is a hard delete — use it when the user asks you to forget something.

Namespaces

The first argument is a namespace, stored as _memory_<namespace>. Use one per person or project to keep memories apart. It must match ^[A-Za-z_][A-Za-z0-9_]*$ — no hyphens.

Good practice

• Store the fact, not the transcript. "Prefers dark roast" beats a paragraph of chat — short, self-contained statements retrieve far better.
• Recall before you answer anything personal, rather than guessing.
• Ask before storing anything sensitive, and never store credentials, payment details, or regulated data.
• Recall before storing to avoid near-duplicates piling up.

Want this to happen automatically?

This skill is manual: you call it when it is relevant. The companion plugin wires the same store into OpenClaw's memory slot so capture and recall happen on their own, and adds SQL-filtered recall, graph relations between memories, and AutoML relevance scoring:

openclaw plugins install clawhub:@synapcores/openclaw-memory

It reads the same SYNAPCORES_API_KEY and talks to the same gateway.

Whether it reads the same memories depends on the engine version. From plugin v0.7.0 on a gateway running v1.14.3-ce or newer, the plugin uses the CREATE MEMORY object — REMEMBER / RECALL / CURRENT / FORGET / TRACE, backed by _mem_<name>_* tables — which is a different store from the _memory_<namespace> table this skill writes to. On older gateways, and whenever the plugin is configured with memory.backend: "legacy", both write the same table and everything here is shared. Check which one is live:

curl -s -H "Authorization: Bearer $SYNAPCORES_API_KEY" "$SYNAPCORES_URL/v1/memory"
# a JSON array  => the plugin will use the CREATE MEMORY object
# HTTP 404      => the plugin stays on the MEMORY_* functions this skill uses

Troubleshooting

Verified against SynapCores v1.14.2-ce.