memv-add-memory

memv-add-memory

Required reading (in order)

1. docs/memv/sdk/memories.md — full method signatures + patterns. Hot path. Read full.
2. docs/memv/core-concepts/memories.md — lifecycle (create → enrich → link → retrieve → decay), what counts as one memory.
3. docs/memv/core-concepts/spaces.md — every write needs space_id. Confirm caller has one.
4. docs/memv/sdk/error-handling.md — exception types to catch.

Hard rules

• space_id mandatory. No global writes.
• Use SDK types (memvai) directly. No wrapper class Memory.
• Push raw content. No pre-summarize text, no pre-extract entities — mem[v] does that.
• Non-text (PDF, image, audio, video) → client.upload.batch.create(...) (skill memv-files / memv-video-ingest), NOT client.memories.add with stringified blob.
• Wrap try/except per sdk/error-handling.md. No bare-except.

SDK surface (verified against docs/memv/sdk/memories.md)

• client.memories.add(space_id, content, metadata?) — write text memory
• client.memories.search(space_id, query, limit?) — retrieve

No update_memory / delete_memory in SDK. Change memory → add new one, let mem[v] graph reconcile. Or client.spaces.delete(space_id=...) to drop whole space.

Skeleton (Python)

from memvai import Memv

client = Memv()  # reads MEMV_API_KEY env var

response = client.memories.add(
    space_id=space_id,                    # required
    content="<raw text>",
    metadata={"source": "...", "user_id": "..."},
)
memory_id = response.memory_id

Verify after writing

Interactive sanity-check write landed:

mcp__memv__search_memory(query="<excerpt>", workspaceId=space_id, maxResults=3)

(workspaceId MCP == space_id SDK — see skill memv-mcp-vs-sdk.)

Don't

• No reuse space_id across logical tenants (per-user/feature/env separation — core-concepts/spaces.md).
• No loop add() thousands of times without backoff. Batch patterns in sdk/advanced.md.
• No store rendered HTML / markup. Push raw source.
• No assume update/delete on client.memories. Don't exist.