Dakera Memory
Persistent, decay-weighted memory for Dify agents, backed by a self-hosted
Dakera server.
Dakera is a self-hosted memory server that adds persistent, decay-weighted vector recall
across sessions: memories are importance-scored and decay over time, so stale context stops
competing with fresh, relevant facts. This plugin lets a Dify Agent, Chatflow, or Workflow
store facts and recall them semantically in later runs.
Tools
| Tool | What it does | Endpoint |
|---|
| Store Memory | Persist a concise, self-contained fact for future recall. Optional importance (0–1), session scope, and tags. | POST /v1/memory/store |
| Recall Memory | Retrieve the most semantically relevant memories for a natural-language query, ranked by score. | POST /v1/memory/recall |
| Search Memory | Filtered browse/list over memories — optional text query plus tag, importance, and count filters. | POST /v1/memory/search |
| Get Memory | Fetch a single memory by its ID (content, importance, tags, metadata). | GET /v1/memory/get/{id} |
| Update Memory | Change an existing memory's content (re-embedded), importance, or tags by ID. | PUT /v1/memory/update/{id} |
| Forget Memory | Delete memories by ID, session, tags, or importance threshold. Requires a selector; deletion is permanent. | POST /v1/memory/forget |
Memories are namespaced by agent_id — use the same agent_id across tools to keep
each agent's (or user's) memories isolated. Recall is best for precise semantic retrieval;
Search is best for browsing/filtering/auditing what an agent remembers.
Setup
1. Run a Dakera server
Dakera is self-hosted. The quickest path is the docker-compose in
dakera-ai/dakera-deploy, which starts the
server (image ghcr.io/dakera-ai/dakera) plus its object store. By default the API listens
on port 3000.
2. Configure the plugin
Provide two credentials when authorizing the tool:
- Dakera Server URL — the base URL of your server, e.g.
http://localhost:3000.
- Dakera API Key (optional) — a
dk-... key if your server was started with
DAKERA_API_KEY. Leave empty for unauthenticated local development.
The plugin validates the connection with a GET /health/live probe when you save credentials.
Usage example
In an Agent app, add the tools you need (all six, or just Store + Recall). A typical loop:
- Early in a task, call Recall Memory with a query like
"user's preferred programming language and coding conventions" (same agent_id you use
everywhere) to pull in relevant prior context.
- When the agent learns something durable, call Store Memory with a concise fact such as
"Alice prefers Rust over Python for backend services", importance = 0.9.
Next session, recalling with the same agent_id surfaces that fact even though the
conversation is new.
Chaining tools with memory IDs
Recall and Search return each memory's ID in their output, so an agent can act on a specific
memory afterwards:
- Recall/Search → Get — pull the full record of a specific hit.
- Recall/Search → Update — correct or re-weight a memory (e.g. bump
importance, replace
content) by its ID.
- Recall/Search → Forget — delete a memory that is now wrong or obsolete by its ID.
Use Forget with a session_id, tags, or below_importance selector to prune a whole set at
once — it refuses to run without at least one selector, so it can't wipe an agent's namespace by
accident.
Development
The plugin source lives alongside the packaged .difypkg in this directory. To run the tests:
pip install dify_plugin requests pytest
pytest -q # exercises all six tools + credential validation against an in-process mock
The tests/ directory is excluded from the packaged plugin via .difyignore.
Requirements & connection
- A reachable, self-hosted Dakera server (this plugin does not bundle or host one).
- Network egress from Dify to the server URL you configure.
- No third-party accounts — all data stays on the server you run.
Links
Contact
Issues and questions: https://github.com/dakera-ai/dakera-py/issues