Migrate from Zep or Graphiti
Move graph-oriented agent memory into FishMem without flattening provenance or history.
Zep and Graphiti integrations often combine conversation episodes, extracted facts, entities, graph edges, and an assembled context block. FishMem exposes a different contract: scoped memory records, source-preserving documents, entities and associations, temporal history, and application-owned prompt assembly. Treat this as a data and behavior migration, not a client rename.
Inventory the live integration
Record the exact SDK and service versions, then list what your application actually reads and writes:
- users, threads, sessions, or groups used as ownership boundaries;
- messages or episodes submitted for extraction;
- facts, entities, edges, and temporal attributes returned to the app;
- context strings, reranking, filters, limits, and latency assumptions;
- deletion, correction, export, and account-erasure paths.
Do not infer the deployed behavior from a current quickstart. Preserve a redacted sample of real request/response traces as the migration acceptance set.
Concept mapping
| Source concept | FishMem target | Migration decision |
|---|---|---|
| user | project plus user_id | Keep the external user identifier; use the project as the application fence. |
| thread or session | run_id or source metadata | Use run_id only when the source boundary has the same lifecycle. |
| message or episode | replay input, memory source, or Document | A transcript is evidence; do not automatically turn every message into a durable memory. |
| extracted fact | Memory | Write infer:false when the source already produced the final record. |
| entity and edge | entity/association projection or metadata | Preserve the source export; rebuild only relationships the target application needs. |
| temporal fact | Memory event time and validity interval | Verify current and historical questions separately. |
| context block | application prompt assembly | Search returns records and evidence; the application decides prompt order and token budget. |
FishMem does not implement the Zep or Graphiti SDKs, graph query language, or context response envelope. It does not claim drop-in compatibility.
Choose the authoritative import
Prefer an official export or a source-system API read over copying a rendered context string. Retain the exported artifact and its creation time. For each imported record, attach enough metadata to find the source again:
await fishmem.memories.add(
{
content: "Alex moved to Berlin in July 2026.",
user_id: "alex",
infer: false,
event_date: "2026-07-01T00:00:00.000Z",
metadata: {
migrated_from: "graph-memory",
migration_batch: "zep-cutover-01",
source_kind: "fact-export",
},
},
{ idempotencyKey: "zep-cutover-01:fact:alex-location" },
);Do not place provider credentials or complete source-system tokens in metadata.
Replay and compare
Run the import in an isolated FishMem project and grade both stored state and downstream behavior:
- Compare counts by user and source kind; inspect rejected and duplicate rows.
- Test current facts, historical facts, corrections, contradictions, and deletion.
- Compare the evidence returned for representative production questions.
- Assemble the prompt with the target application's real token budget.
- Measure error rate, tail latency, and empty or degraded retrieval.
If the source system supplied a preassembled context block, keep it only as a comparison artifact. A successful FishMem migration is not byte-identical context; it is acceptance-set recall that meets the application's decision criteria and remains inspectable.
Cut over with rollback
Dual-write only if the application can give both writes a stable identity and observe partial failure. Otherwise, pause writes for a bounded export/import window. Move one cohort or workspace first, preserve the previous read path, and keep the source export until deletion, recovery, and live recall have all passed.
See Migration rollout for the general release plan and Temporal profile for changing facts.