format.md

The shared storage format

memory-pool and st-qdrant-memory read and write the same Qdrant collections. This is what keeps them compatible. Anything else that follows it can share the pool too, though most interfaces are better off going through the service's HTTP API.

Collections

One collection per agent (st-qdrant-memory keeps one per SillyTavern character). Vectors are cosine-distance embeddings, and every writer must use the same embedding model and dimensions as whatever created the collection (st-qdrant-memory's default is text-embedding-3-large, 3072 dimensions).

Points

Each point is one chunk: a few consecutive messages from one conversation.

Chunk text, dated by its oldest message:

[YYYY-MM-DD]
Speaker: message text
Speaker: message text

Payload:

Field	Type	Meaning
text	string	The chunk text above.
speakers	string	Comma-separated speaker names in the chunk.
messageCount	number	Messages in the chunk.
timestamp	number	Unix ms of the oldest message.
messageIds	string	Comma-separated platform message ids; used to dedup, and to leave out chunks whose messages are already in the reader's context.
isChunk	true	st-qdrant-memory's chunk marker.
source	string	memory-pool only: which interface wrote the point ("discord", "hermes", …). SillyTavern's points have no source; st-qdrant-memory ignores the field. It lets memory-pool's points be filtered and bulk-deleted.
channelId, guildId	string	memory-pool only, optional: where the conversation happened.

Speaker names

For a shared pool to read as one history, the names must match across writers: one name for the agent (selfSpeaker in the service config; when sharing with SillyTavern, its character name there), and one name per person (speakerAliases, keyed by user id; when sharing with SillyTavern, their persona name).

Recall output

POST /retrieve returns the hits plus a formatted block in st-qdrant-memory's style:

[Past chat memories]

• Conversation (Speaker A, Speaker B): "[YYYY-MM-DD] Speaker A: … Speaker B: …" (score: NN%)