Echo Memory
Shared memory for AI agents, as a graph in your own Postgres. Writes never call an LLM.
Open source Repository Open in the app JSON README (API)
About
Shared memory for AI agents, as a graph in your own Postgres. Writes never call an LLM.
Details
- Kind
- MCP servers
- Topic
- AI, RAG & memory
- Publisher
- ayushcodes10
- Origin
- official
- Category
- ferramentas
- Transport
- http
- Version
- 0.3.0
- Stars
- 2
- Forks
- 1
- Last push
- 2026-09-12T22:56:28Z
- Repository state
- ativo
- Language
- Python
- License
- Apache-2.0
- Added
- 2026-09-12 21:02:00
- Updated
- 2026-09-12 21:02:00
- Origin id
io.github.ayushcodes10/echo-mem
README
# Echo Memory
Shared memory for AI coding agents. What Claude Code learns, Codex and Cursor can recall
— in one graph, on your own machine, with every write auditable.
Apache 2.0. The server makes no LLM call on the write path, so storing a memory adds no
inference cost of its own.
## Install
Run it yourself — nothing leaves your machine:
```bash
pipx install echo-mem
echo-memory quickstart
```
That starts the database, applies the schema, and prints the `claude mcp add` command
that registers it with your tools, filled in with the port it actually used. Docker is the only prerequisite; the Postgres image is published, so
nothing is compiled.
Or use the hosted service and run no database at all:
```bash
pipx install echo-mem
echo-memory connect <key> # a key from https://api.echo-mem.com
```
Either way, restart your client afterwards. An MCP server holds the code and config it
started with.
Then, once per machine, so the agent knows *when* to record and recall rather than only
that the tools exist:
```bash
echo-memory install --global
```
## Why
Every AI agent starts from zero unless something remembers what happened last time, and
remembers it well enough and fast enough to still be useful after months or years of
accumulated history. Most memory tools solve short-term recall with plain vector search
over stored facts. That degrades as history grows: more candidates, more noise, slower
retrieval. Echo Memory is built around the read/write algorithm and the data structure
that keeps working at long horizons, not just at day one:
- **A temporal, self-consolidating memory graph.** Facts are edges between entities, not
flat vector rows. Old, rarely-accessed memory doesn't just accumulate: it gets
consolidated into higher-level summaries over time (never deleted, always traceable
back to the original), so retrieval cost stays bounded by what's *currently relevant*,
not by everything that's *ever* been written. See
[`docs/designs/echo-memory-design.md`](docs/designs/echo-memory-design.md#long-horizon-memory-architecture)
for the actual mechanism.
- **Real graph structure, not just similarity.** Multi-hop queries like "how did we end
up here?", answerable because facts are connected, not just individually embedded.
- **Causal typing, not just similarity.** Edges can be tagged `caused_by`, `led_to`,
`blocked_by`, `contradicts`, set by the agent's own read of the conversation, not
inferred statistically. Honest about what's tractable today and what isn't.
- **Auditable by design.** Every change to memory is logged, with a plain-language reason
you can read back (`echo-memory why <fact_id>`). Memory that consolidates and edits
itself is only trustworthy if you can see why.
- **A write path that adds no inference.** Extraction happens in the calling agent, never
on the server, so recording a memory makes zero *additional* LLM calls. The work does
not vanish - it moves to a model that already has the conversation in context - and the
figures below measure the server receiving facts, not the extraction that produced them. Measured locally with
`echo-memory benchmark`: **write 15ms median, query 8ms, digest 1ms, $0.00 inference
cost per episode.** The tradeoff is explicit and worth stating: the agent must arrive
with entities and facts already extracted, which is more work for the caller and the
reason the [MCP tool contract](docs/DEVELOPMENT.md) spells the shape out. The
comparison that makes this matter is Zep/Graphiti, the closest architectural match
(bi-temporal edges, fact invalidation, episode provenance): its own published
description of ingestion is that "every episode triggers multiple LLM calls for
extraction, entity resolution, and invalidation" and that "write cost scales with
volume". Here it doesn't.
- **One storage engine, every scale.** Postgres + pgvector + Apache AGE, from a single
local agent up to an organization-wide shared graph spanning every agent a business
runs. No forced migration later. (The novel work is the memory structure and algorithm
running on top of Postgres, not a new database engine; see the design doc for why.)
- **Any agent, not one vendor's.** The interface is [MCP](https://modelcontextprotocol.io):
any MCP-compatible agent can read and write the same memory graph, whether that's a
coding assistant, a chatbot, an ops agent, or something built in-house.
## Who this is for
- **A developer running local agents** who wants Claude Code, Cursor, or anything else to
stop losing context between sessions and tools.
- **A team or organization running agentic systems in production** (support bots, DevOps
agents, internal tooling) that needs a shared memory layer instead of N disconnected
ones, with the tenancy model (below) to keep it scoped correctly per agent, per team, or
org-wide.
## Status
Early and staged. See [`docs/designs/`](docs/designs/) for the full architecture and the
v1a → v1b build plan. **The validated wedge driving v1a is specifically cross-tool coding
agent memory** (the founder's own daily pain, real and tested). The broader vision above
is the target this architecture is built toward, not yet something v1a itself proves. v1a
proves basic recall works before v1b adds causal typing and multi-hop graph retrieval, and
before v1.1 adds the org-wide tenancy the broader vision depends on.
## Setting it up by hand
`quickstart` is the short way. If you would rather see every step, or you are working on
Echo Memory itself, [`docs/DEVELOPMENT.md`](docs/DEVELOPMENT.md) has the long version:
clone, `docker compose up -d`, `pip install -e ".[dev]"`, `alembic upgrade head`, and the
`claude mcp add` line with its environment.
**Wiring a second tool? Give it its own `ECHO_MEMORY_AGENT_ID`.** Cursor should say
`cursor`, Claude Desktop `claude-desktop`. Memory is shared either way, but a fact records
which tool learned it, and two tools claiming the same id makes cross-tool recall
impossible to see afterwards. `echo-memory adopt` wires every MCP client on the machine at
once, each with its own id, and shows the diff before writing anything.
Scoped to one project instead — a single Claude project, a Cursor workspace, a repo whose
memory should not mingle with the rest? `echo-memory install [path]` writes a
project-scoped MCP config plus a skill (or, for Cursor, an always-applied rule), committed
alongside the code.
See [`docs/INTEGRATIONS.md`](docs/INTEGRATIONS.md) for using Echo Memory from an agent
that does not speak MCP — a chatbot, a DevOps agent, or any custom tool-calling loop.
## The graph
Memory is a graph, not a list of notes. Entities are nodes; a fact is an **edge**
between two of them. That is the whole data model, and everything below follows
from it.
Three projects here. `checkout-api`, `mobile-app` and `data-pipeline` were
recorded in separate sessions and never told about each other, yet the picture
already separates them — because separation is a property of the edges, not a
label anyone applied.
**Clusters come from structure.** Densely connected facts are grouped by label
propagation over the edges, and each cluster is named after its most-connected
node. That is why `data-pipeline` sits apart on the left: nothing it knows
touches payments. It is also why `checkout-api` and `mobile-app` share a cluster
despite being different codebases — they genuinely do share an idea, and the
graph found it rather than being told.
**Components are the stronger claim.** Two nodes in different components have no
path between them at all, which is the strongest statement this graph can make
that two memories are unrelated.
**Projects are a facet, not the structure.** Every fact records the project it
was written from, and you can colour by it, but project says *where a fact was
written*, not *what it belongs with*.
### Click a node: everything it takes part in
`idempotency keys` is the concept that joined those two codebases. The panel
shows it referenced from **checkout-api twice and mobile-app once**, the three
facts it appears in, and how the node itself resolved — each mention matched an
existing node by exact name rather than creating a duplicate.
Nobody wrote "these projects are related." Two sessions independently recorded a
fact about idempotency keys, entity resolution matched them to one node, and the
relationship exists as a consequence.
### Click a link: why memory believes it
This is what a knowledge graph gives you that a code map cannot. Selecting the
edge answers, for that single fact:
| | |
|---|---|
| **what** | the sentence, its `relation_type`, and how confidently it was stated |
| **when** | when it became valid, and when it was superseded if it has been |
| **who** | which agent wrote it, in which session |
| **where** | which project it came from |
| **why** | the audit trail — created, superseded from what to what, and the entity-resolution rationale for the nodes at either end |
A superseded fact is never deleted. It stops being drawn, because the graph no
longer asserts that relationship, but it stays reachable from its node and keeps
its full history. `echo-memory why <fact_id>` prints the same trail in a terminal.
### Seeing your own
```bash
echo-memory dashboard --serve --open
```
The images above come from a synthetic dataset (`scripts/demo-seed.py`) rather
than a real store, for the obvious reason: a real memory graph is full of
hostnames, account numbers and client names.
## Architecture
- **Storage:** PostgreSQL with the `pgvector` and Apache AGE extensions
- **Retrieval:** hybrid vector + full-text search (v1a), with Personalized PageRank via
`networkx` added in v1b for multi-hop associative retrieval
- **Interface:** [Model Context Protocol](https://modelcontextprotocol.io) server:
`write_episode`, `query_memory`, `record_recall_save`, `get_audit_log`
## Is the graph in good shape?
```
echo-memory health
```
A score, what is strong, what needs attention, and what to do about each,
including what recall has cost: how often memory was read, how often a read
returned anything, roughly how many tokens were injected, and how many saves
those reads produced. Writes were counted from the start; reads were not counted
at all, so nothing could answer whether recall earns what it costs. It
exists to be run when you have no question - a store can look healthy by every
number this CLI reports while most of its facts came from a bulk import, the
last real write was a week ago, and only one of several wired agents has ever
written anything. `--json` for machine-readable output.
Nothing in it is gated. The paid tiers sell hosting and the things that only
exist when several people share a graph; diagnostics about your own data are not
a thing to withhold from the person whose data it is.
## Contributing
See [`CONTRIBUTING.md`](CONTRIBUTING.md). Issues and PRs welcome; please read the design
docs first so proposals fit the staged build plan. A first pull request is asked to sign
the [Contributor License Agreement](CLA.md) — once, in the PR thread.
## License
Apache License 2.0. See [`LICENSE`](LICENSE).
<!-- The MCP registry proves you own a PyPI package by finding this line in the
package's own description. It has to survive into the built distribution,
which is why it lives in the README rather than in a workflow. -->
mcp-name: io.github.ayushcodes10/echo-mem