Echo Memory

Echo Memory

Shared memory for AI agents, as a graph in your own database. What Claude Code learns, Cursor and Codex can recall. Every fact records who wrote it and when, and the server never calls a model to store one.

Every line above is real output, captured from these commands against a throwaway database. The transcript is at .github/branding/demo-transcript.txt.

Your agents start every session from zero. The usual fix is a notes file you paste into context, which grows until it is mostly irrelevant to whatever you are asking. Echo Memory is the other shape: facts connected to each other, and a query that returns the few that matter. On the author's own 325 fact store that is 96.7% less context for the same answer, with the answer still present 87.2% of the time across 1,190 queries the harness derived from their own answers.

write_episode                          query_memory
  billing ──uses──▸ Razorpay             "how do we take payments"
    │ written by claude-code               ▸ billing uses Razorpay, not Stripe
    │ supersedes ──▸ Stripe                  written by claude-code, 3 days ago
    └ no model invoked                     ▸ 1,372 tokens, not 41,838

Install

Requires: Python 3.11+ and Docker (for the database).

pipx install echo-mem
echo-memory quickstart

quickstart starts the database, applies the schema, and prints the claude mcp add line that registers it, filled in with the port it actually used. The Postgres image is published, so nothing compiles.

Or use the hosted service and run no database at all:

pipx install echo-mem
echo-memory connect <key>          # a key from https://app.echo-mem.com

Then once per machine, so an agent knows when to record and recall rather than only that the tools exist:

echo-memory install --global

Restart your client afterwards. An MCP server is a long lived process that holds the code and config it started with, and an editable install does not change that.

The PyPI name is echo-mem, not echo-memory. That name belongs to an unrelated hosted product. The import package and the CLI are both echo_memory / echo-memory; only the distribution name differs.

Upgrading

pipx upgrade echo-mem
echo-memory init-db                # apply any new migrations
echo-memory reindex                # rebuild what the new code derives

Both steps, in that order, and neither is optional on a store that predates the version you just installed.

Skipping init-db fails loudly. A read against a schema older than the code raises an undefined column, which is the good case: it stops rather than answering from something it half understands.

Skipping reindex fails silently, which is worse. Derived data that the new code reads differently is still the old data, and nothing errors. Concretely, on an upgrade to 0.5.3 the lexical corpus statistics are treated as stale until rebuilt, so BM25 stays off and your queries keep ranking by ts_rank while the setting reports as on. reindex rebuilds the embeddings and those statistics together. echo-memory health tells you afterwards.

Usage

echo-memory status                 # what each scope holds, and which agents have written
echo-memory health                 # a score, what is weak, and what to do about it
echo-memory dashboard --serve      # the graph, in a browser, localhost only

echo-memory why <fact_id>          # the full audit trail for one fact
echo-memory recall "<question>"    # query the store from a terminal
echo-memory export                 # everything, as JSON

echo-memory install --for cursor   # wire one client, project scoped
echo-memory adopt                  # wire every MCP client on the machine, each with its own id

echo-memory infer-causal-hints     # type the facts whose own sentence states a cause (dry run)

echo-memory eval                   # retrieval quality against your own store
echo-memory eval --context         # what a recall costs against injecting everything
echo-memory eval --context --sweep # the same, as a curve across corpus size
echo-memory eval-external locomo <path>       # LoCoMo, the corpus the published figures use
echo-memory eval-external longmemeval <path>  # LongMemEval, the same
echo-memory calibrate              # is entity resolution trustworthy on your data
echo-memory benchmark              # write, query and digest latency

The seven MCP tools

What you get

A graph, not a list. Entities are nodes and a fact is an edge between two of them. Two sessions that never knew about each other resolve onto the same entity by name, so the second inherits what the first learned.

Bounded retrieval, designed but not built. The plan is that old, rarely read memory demotes into higher level summaries over time, with nothing discarded and every summary still edged back to the facts it came from. None of it exists yet: there is no tiering, no summarisation, and retrieval today walks every active fact in the scope. It is described in docs/designs/ and listed below as v1c, and this paragraph used to claim it in the present tense.

A ranked answer that says why each fact is in it. Every returned fact carries score, its cosine similarity to the query, rank, its position, and matched, the channels that found it. score is deliberately not the fusion number: reciprocal rank values are sums of 1/(k+rank) and mean nothing from one query to the next, while a similarity means the same thing every time, which is what a caller thresholding on it needs. It is computed for every fact returned, including the ones only full text search found, because which channel retrieved a fact is an implementation detail and has no business reaching a field callers read as relevance.

The full text channel ranks by BM25. Postgres's ts_rank counts term occurrences and nothing else: no inverse document frequency, so a word in every fact of a scope counts as much as one in three; no saturation, so repetition scales linearly; no length normalisation, so a long fact is punished for being long. On 324 real prompts the rank 3 and rank 4 scores were identical in 59% of them, and reciprocal rank fusion reads rank position and never the score underneath it, so a channel ordered by scan order handed the fusion a coin flip.

BM25 is computed in SQL rather than installed, so no extension is added to your database. Corpus statistics are rebuilt amortised on write inside the transaction that already holds the scope's lock, and by echo-memory reindex; a scope that has none falls back to ts_rank rather than ranking on statistics it does not have. ECHO_MEMORY_LEXICAL_BM25=0 returns to ts_rank everywhere.

Facts about an entity, which resemblance cannot express. query_memory takes about, one or two entity names, and returns only the facts recorded against them. One name gives the edges incident to that node; two give the edges between them in either direction, which is "both names in the same fact" without going near the fact text.

This is not a ranking improvement and it is not a substitute for one. A query about one team returns a semantically close fact about a different team because the wrong fact genuinely does resemble the question, and no amount of better ranking removes it. Identity is not similarity, and a fact is an edge between two nodes, so asking which entity a fact is about is structural and exact.

Name matching is exact and case insensitive, never a substring, because a short name is a substring of longer unrelated words. A name nothing is recorded under returns no facts rather than the nearest thing, and so does a pair with no fact joining them: that is real absence and it means nobody recorded this, not here is something adjacent. More than two names is refused rather than answered empty, because an edge has two endpoints and a fact about three entities is a question this model cannot express.

The history the store has always kept is readable. Every fact has carried t_valid and t_invalid since the first migration, and the read path had only ever asked whether a fact is current, so the store paid to keep the whole record of what a scope believed and could not answer the first question a post mortem asks. query_memory takes as_of in unix seconds and every channel honours it: vector, full text, the graph hop, the digest.

Supersession is what makes that more than a curiosity. Writing the same source, target and relation again with a changed claim does not edit the old fact, it ends it: the old edge takes a t_invalid and stays queryable. Nothing is ever rewritten, so the history is real rather than reconstructed, which is the property the memory reconsolidation literature gives up and the reason this project refuses that idea.

Restating an unchanged claim is not a change and no longer counts as one. If the fact text, the confidence and the causal hint all match the edge already there, the write is skipped and reported under restated. It used to create a second edge, end the first, and emit a supersession whose before_fact equalled its after_fact. That is the mechanism behind one customer edge carrying 47 versions of a single sentence, and behind 543 writes producing 112 facts while every one of them was billed. It also makes a retry safe: with no idempotency key, a caller retrying a write that had actually landed used to burn a version on it.

Every write answers whether it stored anything. write_episode can end with an empty edges_created four different ways, and from the caller's side they used to be indistinguishable without knowing which other key to inspect. Each one now reports stored: false and names itself in not_stored: refused (validation rejected it), deferred (an entity mention was close enough to an existing node to be worth asking about and not close enough to merge), unchanged (every fact already existed word for word) and no_facts (the episode asserted none). A refusal carries the same keys as a success, so one check works on every path, and retryable separates a database blip worth retrying from a validation error that will fail forever. A write that stored some of its facts and held the rest says partial.

The deferral is the one that cost real data, because it looks most like success: a well formed response, no error, isError false, and nothing written. A customer loading 26,633 facts read that as 26,633 successes and stored 1,650.

Provenance on every fact. Who wrote it, which tool, which project, when, and which reads returned it. A superseded fact is never deleted. It stops being drawn and stays reachable with its history.

Causal typing, and a walk over it. A fact can carry causal_hint: one of caused_by, led_to, enabled_by, blocked_by or contradicts, set by the agent's own read of what the session said and never inferred statistically. trace_cause walks those links and returns chains rather than a ranked list, because the answer to "why did this happen" is an ordered chain and similarity cannot produce one. "A led_to B" and "B caused_by A" are one claim written from opposite ends, and both assemble into the same chain.

A fact with no hint is associative, which is the default and usually correct. An empty answer from trace_cause says which kind of empty it is: nobody recorded a cause here is a different fact about a store than there are no chains.

Filling it in, for a store that predates it. Facts written before 0.5.0 carry no hint, so on an existing store trace_cause has nothing to walk. echo-memory infer-causal-hints re-reads the fact text already stored with your own model (ECHO_MEMORY_LLM_API_KEY, ECHO_MEMORY_LLM_MODEL) and types the edges whose own sentence states the relation. Dry run by default; --write applies what it printed and --clear --write takes it back.

This is extraction done late, not causal discovery, and the difference is enforced rather than promised. A sentence with no causal connective is never sent to a model, so co-occurrence is refused before it costs anything. Every proposal has to quote the words that state the relation, and a quote that is not literally in the fact is dropped, so a model reasoning from the world instead of reading the sentence gets nothing stored. Each hint written is audited and marked as extracted late, which is what makes --clear able to remove exactly these and never one you wrote at write time.

No inference on the write path. Extraction happens in the calling agent, so storing a memory invokes no model on the server. Three commands in the engine reach a provider and all three are worth naming rather than glossing: infer-causal-hints above, distil, and judge-external, which grades a finished benchmark run. None is an exception to the claim. Each is a command a person types against facts already stored, none runs as part of a write, a migration or a hook, and src/echo_memory/cli/llm.py holds the only outbound request in the engine, so a store that runs none of them never causes a model call. The cost moved rather than vanished: the agent has to arrive with entities and facts already extracted, which is what the tool contract spells out. The comparison that makes this matter is Zep/Graphiti, the closest architectural match, whose own description of ingestion is that "every episode triggers multiple LLM calls" and that "write cost scales with volume".

The obvious reply is that cheap writes are cheap because they do less, and that reply is correct on the mechanism. docs/WRITE-COST.md answers it properly, including the two measured costs of the choice: the Stop gate fired seven times and produced one fact, and a write touching an ambiguous entity is deferred rather than stored. The deferral is still the cost; what has changed is that the call now says so instead of returning as though it had succeeded.

Any MCP client. A coding assistant, a chatbot, an ops agent, or something built in house. Coding agents are where this is proven, not what it is limited to.

Numbers, and how they were taken

Every figure comes from this repository or a live store, with the command that reproduces it on yours and the size of the store it was taken against. The corpus is small and the noise floor is stated, because a difference nobody sized is not a result.

Two of these describe the store rather than the product. eval --context and eval --ablate build their queries from the facts they are scoring, so the labels are free and the absolute values belong to this store alone; they are honest about a difference between configurations and say nothing about how echo-mem would rank on somebody else's data. The two external rows are the ones to compare against anything.

That last row is the one that went the wrong way, and it is here on purpose. The interval includes chance, so the unattended merge is switched off: at the automatic bar precision was 50% over two reviewed pairs, and the audit log showed that path had fired once in the system's entire history. A near match is now offered for confirmation instead.

The LoCoMo row is retrieval, not QA accuracy. Published LoCoMo results have a model write an answer and a second model judge it; this asks only whether the turn holding the answer came back, which is a ceiling on QA accuracy rather than a substitute for it, and is not comparable to anybody's published QA figure. It also feeds raw dialogue turns, which skips the extraction step this design pushes to the calling agent, so it is a floor as well as a ceiling. The worst row, multi hop at recall@1 0.099, is in docs/BENCHMARKS.md with the rest.

The context saving is measured against a specific baseline, stated so it cannot be read as more than it is. Not "no memory at all", which is however long a human spends re explaining and is unmeasurable. It is the thing people do instead: keep the project's notes in one file and paste the whole file. On that store the file is 325 facts, about 41,838 tokens; a recall returned 1,372 on average. The hit rate belongs beside it, because a recall that returned nothing would score 100%.

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 else 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 rather than 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: nothing it knows touches payments. It is also why checkout-api and mobile-app share a cluster despite being different codebases. They genuinely 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.

Click a node: everything it takes part in

idempotency keys is the largest node here and nobody made it large: seventeen facts from several services resolved onto one entity by name. The panel lists every one, with which agent wrote it and when.

Click a link: why memory believes it

Not a tooltip. Who wrote the fact, in which project, when, and how each of its entities resolved. echo-memory why <fact_id> prints the same trail in a terminal.

Seeing your own

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.

Wiring more than one tool

Give each client 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                  # every MCP client on the machine, each with its own id
echo-memory install [path]         # one project: MCP config plus a skill, committed with the code

adopt shows the diff before writing anything. For an agent that does not speak MCP, see docs/INTEGRATIONS.md.

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, because a store can look healthy by every other number 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 plan sells hosting; diagnostics about your own data are not a thing to withhold from the person whose data it is.

Architecture

Storage PostgreSQL with pgvector and Apache AGE, from a single local agent up to an organisation wide shared graph, with no forced migration later. The novel work is the memory structure and the read/write algorithm on top of it, not a new database engine.

Retrieval Hybrid vector and full text search fused by reciprocal rank in v1a, with the full text side ranked by BM25 computed in SQL over per scope corpus statistics. Personalised PageRank via networkx lands in v1b for multi hop associative retrieval.

Interface Model Context Protocol, so any compliant agent reads and writes the same graph.

Status

Early and staged, on purpose. See docs/designs/ for the architecture and the v1a to v1b plan.

The validated wedge driving v1a is memory shared across coding agents, which is the author's own daily pain and the case with the most evidence behind it. Everything else is the target this architecture is built toward.

Hosted

app.echo-mem.com runs the database for you: $149 a month for 1,500,000 units, where a unit is one fact written or one fact returned. It is paid from the first key, because hosting is the thing it charges for, and it runs the same published package you would install yourself.

Running it yourself is free for development, testing and evaluation at any size, and free in production for organisations under 50 people and under $5M revenue, with no account and no feature held back. Above either threshold, production use needs an agreement: talk to us.

See CONTRIBUTING.md. Issues and pull requests welcome; please read the design docs first so proposals fit the staged build plan. Commits need a sign-off line, git commit -s, which is the Developer Certificate of Origin. There is no contributor agreement to read.

The most useful contribution is a measurement that disagrees with one of the numbers above. Run echo-memory eval, calibrate or benchmark on your own store and open an issue with the output.

License

Apache 2.0. See LICENSE. Run it, fork it, build on it, ship it in your product, at any size, with no agreement and nobody to ask.

For about a month this was not true. Versions 0.5.0 to 0.5.5 shipped under the Business Source License, which made production use free only below fifty people and $5M of revenue. That is withdrawn. Ateleir Tech holds the copyright in the whole work and relicenses all of it, those versions included, so the grant only ever got wider and nobody lost a right they already had.

The reasoning, since a project that changes its licence twice owes you one: the restriction was protecting the wrong thing. What makes this hard to run is not the source, it is a Postgres with Apache AGE and pgvector that somebody has to operate, back up and upgrade. The hosted plan sells exactly that and is unaffected by the licence. Meanwhile the restriction was costing the thing the project most needs, which is for people to be able to use it without reading a grant first.

What is not Apache 2.0 is the hosted service itself: the metering, billing and tenancy that only exist because somebody is being invoiced. None of it sits in the read or write path.

mcp-name: com.echo-mem/echo-mem