recall-echo
Your coding agent forgets everything the moment you close the session.
So you explain it again. The architecture. Why you dropped that library. The deployment quirk that bites every time. The convention you agreed on last Tuesday. You've written it down in a file, and the file is now 400 lines, and it's stale, and you're still explaining.
recall-echo gives your agent an actual memory: it remembers what you told it, gets surer about what's true as you keep working, and can be asked about any of it — without you maintaining a thing.
|
That's it. init finds the agent CLIs you already have, picks one to extract with (asking only if there's a real choice), installs Claude Code's hooks, registers the MCP server with every client it found, and downloads the embedding model up front. Conversations are then captured when a session ends, turned into knowledge while your machine is idle, and available to your agent the next time it needs them.
What makes it different
You don't curate it. Archiving, checkpointing and extraction are automatic. Memory that depends on the agent remembering to save things is circular — this closes that loop.
It finds things by meaning. Ask about "authentication" and it surfaces the session about JWT and login flows, even if nobody used that word.
It gets surer over time. Every fact carries confidence that climbs as conversations corroborate it and decays when they don't. Stale claims lose weight on their own. One offhand contradiction won't erase something the graph is confident about — overturning that takes sustained evidence.
It knows who said what. What you stated outweighs what the agent inferred from its own notes, and an agent repeating itself is counted as repetition, not proof. Memory can't drift into an echo chamber of its own output.
It runs on your machine. Embedded database, local embeddings, nothing leaves the box. And it runs on whatever you already pay for — Claude, Grok, Codex, Gemini, or fully local with Ollama. Retrieval costs nothing at all; only learning new things needs a model.
Your agent can ask it questions. It speaks MCP, so Claude Code, Codex, Grok, Gemini, Cursor or Zed can query memory mid-conversation, in its own words. init registers it with every one of those CLIs it finds, so there is nothing to wire up.
Does it work
On a LongMemEval subset it went from answering 0 of 9 questions correctly to 7 of 9, with the right evidence retrieved for every single one. That's a small sample and it's stated as such — the methodology, the regressions we hit along the way, and the honest limitations are all in docs/benchmarks/.
Using it
Day to day, you don't. That's the point — sessions are captured and turned into knowledge without you doing anything.
When you do want to poke at it:
And from inside your agent, once the MCP server is registered, it asks for itself:
You: why did we drop the websocket approach? Agent: (calls
recall_query) You moved off it in June — the reconnect logic kept dropping messages under load, and you settled on polling with a 30s interval.
No prompting required; the agent decides when it needs to remember something.
Architecture
recall-echo provides a four-layer memory model:
┌──────────────────────────────────────────────────────────┐
│ MEMORY ARCHITECTURE │
│ │
│ Layer 0: KNOWLEDGE GRAPH (structured, semantic) │
│ ┌──────────────────────────────────────────────────┐ │
│ │ SurrealDB + FastEmbed │ │
│ │ Entities, relationships, episodes │ │
│ │ Bayesian confidence · Semantic search (HNSW) │ │
│ │ LLM-powered extraction + deduplication │ │
│ └──────────────────────────────────────────────────┘ │
│ │
│ Layer 1: CURATED (always in context) │
│ ┌───────────┐ │
│ │ MEMORY.md │ Facts, preferences, patterns │
│ └───────────┘ Distilled & maintained by the agent │
│ │
│ Layer 2: SHORT-TERM (FIFO rolling window) │
│ ┌───────────────┐ │
│ │ EPHEMERAL.md │ Last N session summaries │
│ └───────────────┘ Appended on archive, auto-trimmed │
│ │
│ Layer 3: LONG-TERM (searched on demand) │
│ ┌─────────────┐ ┌────────────────────────────┐ │
│ │ ARCHIVE.md │───→│ conversations/ │ │
│ └─────────────┘ │ conversation-001.md │ │
│ │ conversation-002.md │ │
│ │ ... │ │
│ └────────────────────────────┘ │
│ YAML frontmatter + markdown │
│ LLM-summarized or algorithmic │
└──────────────────────────────────────────────────────────┘
Knowledge Graph (Layer 0, default)
The knowledge graph is the structural foundation of recall-echo. It turns conversation archives into structured, searchable memory. Enabled by default via the graph feature.
What it does. When conversations are archived, recall-echo extracts entities (people, projects, tools, concepts) and the relationships between them, then stores them in an embedded SurrealDB graph database. Semantic search via fastembed embeddings lets agents find relevant memories by meaning, not just keywords — so a search for "authentication" surfaces conversations about JWT, OAuth, and login flows even if those exact words weren't in the query.
Why Bayesian confidence. Traditional knowledge graphs store facts as absolutes — "Dani uses NeoVim" is either true or not. But memories aren't binary. Things change, context matters, and some things are more certain than others. recall-echo uses a Beta-Binomial Bayesian confidence model on every relationship edge:
- Each relationship starts with a confidence prior based on how it was established: authoritative (1.0), explicit (0.9), inferred (0.6), or speculative (0.3)
- Evidence is persisted on the edge as Beta pseudo-counts (
alphafor corroboration,betafor contradiction). Corroboration adds to α, contradiction adds to β, and the stored confidence is the posterior meanα / (α + β) - Because the counts accumulate, the graph distinguishes "believed at 0.9" from "believed at 0.9 for good reason" — the posterior variance narrows as evidence builds, and a prior's initial skepticism persists rather than being recomputed away
- Updates are gradual — a prior is worth ~10 observations, so a single contradictory mention doesn't erase established knowledge
- Observations are weighted by provenance. Every episode is stamped at ingestion with who authored it, and an observation contributes evidence accordingly: an independent external source counts fully (1.0), the human nearly fully (0.8), the agent restating its own belief almost not at all (0.05, configurable via
[graph.provenance]). Self-corroboration is also tallied separately on the edge, so coherence never passes for evidence - Multi-hop queries compound confidence along the path, naturally preferring shorter, higher-confidence routes
This means the graph handles contradictions, reinforces patterns over time, and lets uncertain or stale knowledge fade gracefully — instead of requiring manual cleanup or producing false-positive retrievals.
Entity types: person, project, tool, service, preference, decision, event, concept, case, pattern, thread, thought, question, observation, policy, measurement, outcome. Mutable types (person, project, tool, etc.) can be updated; immutable types (decision, event, case, etc.) are append-only.
Extraction pipeline: When conversations are archived, an LLM-powered pipeline chunks the text (~500 tokens), extracts entities and relationships in parallel (up to 10 concurrent), then deduplicates sequentially. Dedup escalates to the LLM only for candidates it cannot settle itself: an existing entity of the same name and type is the same entity, a nearest neighbour above certain_similarity is the same entity, one below review_similarity is a new one, and only the band in between buys an LLM skip/create/merge decision — over a set capped at max_candidates. The bands are cut on raw cosine similarity, never on the blended retrieval score, so a popular entity does not read as a likelier duplicate and dedup cost stays flat as the graph grows. Re-extracted relationships receive Bayesian corroboration updates weighted by the provenance of the chunk they came from — conversation turn roles are read to tell the human's words from the agent's, and graph ingest --external marks genuinely external material — so knowledge confirmed by independent sources gains confidence while the agent repeating itself barely moves the score.
Tiered content: Entities store content at three levels — L0 (abstract, used for embeddings and cheap traversal), L1 (overview, used for reranking), and L2 (full content, pulled on demand). This keeps graph traversal fast.
Graph commands:
# Core
# Search & traversal
# Data management
# Pipeline & integrations
All paths are relative to an entity root directory:
{entity_root}/memory/
├── MEMORY.md # Layer 1 — curated facts (≤200 lines)
├── EPHEMERAL.md # Layer 2 — rolling session window (default 5)
├── ARCHIVE.md # Layer 3 — conversation index
├── conversations/ # Layer 3 — full conversation archives
│ ├── conversation-001.md
│ ├── conversation-002.md
│ └── ...
├── graph/ # Layer 0 — knowledge graph
│ ├── surreal/ # SurrealDB embedded data
│ └── models/ # FastEmbed cached models
└── .recall-echo.toml # Optional configuration
How It Works
recall-echo operates in two modes:
As a pulse-null Plugin
recall-echo is a native pulse-null plugin implementing the Plugin trait from pulse-system-types. It fills the required Memory role (exactly one per entity).
- pulse-null calls
archive::archive_session()at session end — creates a conversation archive with LLM-generated summary, updates ARCHIVE.md index, appends to EPHEMERAL.md - pulse-null calls
checkpoint::create_checkpoint()before context compaction — preserves conversation state before details are lost - Health checks report memory directory state (Healthy / Degraded / Down)
- Setup wizard prompts for entity_root during
pulse-null init
use RecallEcho;
// pulse-null creates the plugin via factory:
let plugin = create.await?;
// plugin.role() == PluginRole::Memory
As a Standalone CLI
For administration and use outside pulse-null:
Installation
cargo install
--locked installs the exact dependency versions the release was tested
against. The embedded graph store's on-disk record format is tied to the
SurrealDB version, so a build that resolves a different SurrealDB may be
unable to read a store another build wrote.
Prebuilt binaries
Download from GitHub Releases for:
x86_64-unknown-linux-gnuaarch64-unknown-linux-gnux86_64-apple-darwinaarch64-apple-darwin(Apple Silicon)
From source
Commands
recall-echo init
Create the memory directory structure under entity_root. Creates memory/ with MEMORY.md, EPHEMERAL.md, ARCHIVE.md, and conversations/. Idempotent — never overwrites existing files.
recall-echo status
Health check with a dashboard showing memory usage, ephemeral state, archive count, recent sessions, and health assessment. Color-coded bars show MEMORY.md capacity (green → yellow → red at 75% / 90%).
recall-echo — healthy
MEMORY.md: 142/200 lines (71%)
EPHEMERAL.md: 3 entries
Archives: 23 conversations
recall-echo search
Search conversation archives.
Ranked search scores files by match count, word coverage, and recency.
recall-echo distill
Analyze MEMORY.md and suggest cleanup. Identifies sections over 30 lines that could be extracted to topic files (e.g., memory/debugging.md) with references left in MEMORY.md.
recall-echo consume
Output EPHEMERAL.md content wrapped in memory markers. Used by hooks or scripts that need to inject recent session context into an agent's input.
recall-echo archive-session
Archive a Claude Code session from a JSONL transcript. Extracts messages, generates a summary (LLM-powered when available, algorithmic fallback), updates ARCHIVE.md, and appends to EPHEMERAL.md. Designed to run as a SessionEnd hook.
recall-echo checkpoint
Save a checkpoint before context compression. Creates a numbered checkpoint file so the agent can fill in summary details. Designed to run as a PreCompact hook.
recall-echo graph
Knowledge graph operations. See the Architecture section above for the full command list.
Search & traversal:
graph search <query>— Semantic search across entities. Supports--limit,--type(filter by entity type), and--keyword(filter by name/abstract).graph query <query>— Hybrid query combining semantic search, confidence-weighted graph expansion, and optional episode retrieval. Supports--depth(expansion depth, default 1, 0 = semantic only),--episodes(include episode results),--limit,--type,--keyword.graph traverse <entity>— DFS traversal from a named entity with cycle detection. Displays confidence percentages on edges (e.g.[85%]). Edges below 0.1 confidence are filtered. Supports--depth(default 2) and--type-filter.
Data management:
graph add-entity— Manually add an entity. Requires--name,--type,--abstract. Supports--overviewand--source.graph relate <from> --rel <type> --target <to>— Create a relationship between two entities. Supports--descriptionand--source.graph ingest <archive>— Ingest a single archive file (creates episodes, no LLM required).graph ingest-all— Scan conversations/ and ingest all un-ingested archives.graph extract— LLM-powered entity extraction. Supports--log <N>(single archive),--all(all un-extracted),--dry-run,--model,--provider(any name from LLM providers),--delay-ms. The daemon runs this pass on its own once the machine is quiet (see Background extraction); this command is how you run it now, or inservermode, or after changing the model.
Daemon:
graph daemon status— Socket path, pid, version and uptime of the daemon serving this graph.graph daemon stop— Stop that daemon. The next graph command starts a fresh one.
Pipeline & integrations:
graph pipeline sync— Sync pipeline documents (LEARNING.md, THOUGHTS.md, CURIOSITY.md, REFLECTIONS.md, PRAXIS.md) into the graph. Idempotent — diffs parsed entries vs existing graph entities.graph pipeline status— Pipeline health with staleness tracking.graph pipeline flow <entity>— Trace an entity's lineage through the pipeline stages.graph pipeline stale— List stale pipeline entities. Supports--days(threshold, default 7).graph vigil-sync— Sync vigil-pulse metacognitive signals and caliber outcomes into the graph as Measurement and Outcome entities. Supports--signals-pathand--outcomes-path.
recall-echo serve
Runs the graph daemon for a memory directory. You never need to run this by hand — graph commands and hooks start it automatically. It exists for supervised deployments:
--foreground logs to stderr as well as <memory_dir>/graph/daemon.log and
disables idle shutdown, leaving lifetime to systemd. Background extraction
still runs — it waits for quiet, not for an idle timeout.
recall-echo mcp
An MCP server over stdio, so an agent can query its own memory mid-conversation.
Why it exists. Without it the knowledge graph is effectively write-only.
SessionEnd ingests episodes and SessionStart runs consume, which only
prints EPHEMERAL.md — nothing in a normal session ever reads the graph, so the
Bayesian confidence, semantic search, provenance weighting and temporal decay
sit behind a command a human has to type by hand. The MCP server is the read
path: the agent asks memory the actual question, at the moment it matters.
You don't have to register it. recall-echo init does that for every agent
CLI on the machine. To add it by hand, or to a client init doesn't know, each
vendor spells it differently:
Or, equivalently, in a project's .mcp.json:
--entity-root defaults to the current directory, so it can be omitted when
the client is launched from the entity root.
Tools. All five are read-only; none can write to the graph.
| Tool | Answers |
|---|---|
recall_query |
The default lookup: semantic search + one hop of graph expansion + the conversation fragments behind it |
recall_search |
Semantic entity search alone — names, types, abstracts, retrieval scores |
recall_episodes |
The raw conversation fragments, for what was actually said |
recall_traverse |
Relationships out of one named entity, as a tree with edge confidence |
recall_status |
Entity, relationship and episode counts — tells an empty memory from a failed lookup |
Every tool runs through the same graph daemon as the CLI, so it inherits the daemon's auto-start, locking and concurrency, and starting an MCP client never takes the store away from a hook.
Writing is deliberately absent. The graph discounts what the agent asserts
about itself ([graph.provenance]), and a tool that let the model create
entities and edges directly would route around exactly that mechanism. Memory
is written on the ingest path, where every episode is stamped with its
authorship.
Archive Format
Conversation archives use YAML frontmatter with markdown content:
---
log: 5
date: "2026-03-06T10:30:00Z"
session_id: "abc123"
message_count: 34
duration: "30m"
source: "session"
topics:
---
## Summary
Summary of the conversation with key outcomes.
**Decisions**: Chose JWT for authentication.
**Action Items**: Implement token refresh endpoint.
### User
(message content)
### Assistant
(message content)
## Tags
**Files**: src/auth.rs, src/middleware.rs
**Tools**: Read, Edit, Bash
Summaries are LLM-generated when a provider is available (via pulse-null), with silent fallback to algorithmic extraction.
LLM providers
Entity extraction and dedup need a model. recall-echo will use whichever agent CLI you already pay a subscription for, or a local model, or an API key — your choice, one config key. Embeddings are always local (fastembed/ONNX, no network after the model downloads once), so semantic search, HNSW indexing and graph traversal never cost anything regardless of this setting.
provider |
How it runs | What it costs | Verified |
|---|---|---|---|
claude-code |
spawns claude |
Claude subscription — no per-token billing | yes, live call (claude 2.1.x) |
gemini |
spawns gemini |
Gemini subscription/free tier — no per-token billing | yes (gemini 0.27.x) |
grok |
spawns grok |
Grok subscription — no per-token billing | yes, live call (grok, JSON envelope) |
codex |
spawns codex exec |
ChatGPT/Codex subscription — no per-token billing | yes, live call (codex-cli 0.146.x, NDJSON stream) |
cli |
spawns whatever [llm.cli] describes |
whatever that CLI costs | n/a — you supply the flags |
ollama |
HTTP to localhost:11434/v1 |
free, fully local | yes |
openai |
HTTP to any OpenAI-compatible endpoint | per token (API key) | yes |
anthropic |
HTTP to the Anthropic API | per token (API key) | yes |
The five CLI providers are one implementation. A provider name selects a
preset — a set of defaults for the keys in [llm.cli] — and every key can be
overridden, so a CLI with no preset is configuration rather than a new release:
[]
= "cli"
[]
= "mycli" # binary name or path
= ["chat"] # fixed args before the generated flags
= "flag" # "stdin" | "flag" | "arg"
= "--ask" # used when prompt_delivery = "flag"
= "--model" # omitted when empty, or when no model is set
= "--format" # passed alone when output_format_value is empty
= "json"
= "" # empty prepends the system prompt to the message
= "single-json" # "raw" | "single-json" | "ndjson"
= "data.text" # dotted path; "" or omitted = stdout is the answer
= [] # line selectors, ndjson mode only
= ["--quiet"]
= 300 # 0 waits forever
Output handling is deliberately forgiving: if the JSON cannot be parsed, or the path is not there, or it holds something other than a string, the CLI's stdout is used verbatim rather than failing the call. A non-zero exit is an error carrying the CLI's own stderr, and so is a JSON envelope that reports its own failure while exiting zero.
Presets, exactly as they are called (<prompt> is the message,
<system> the system prompt; < means stdin):
claude-code claude -p --model <M> --output-format text --system-prompt <system> \
--no-session-persistence < <prompt>
gemini gemini -m <M> -o json -p "<system>\n\n<prompt>"
grok grok -m <M> --output-format json -p "<system>\n\n<prompt>"
codex codex exec -m <M> --json --skip-git-repo-check < "<system>\n\n<prompt>"
Only claude takes a system prompt as a flag; for the others it is prepended to
the message. -m is omitted entirely when no model is configured, so each CLI
uses its own default. Set CLAUDE_BIN, GEMINI_BIN, GROK_BIN, CODEX_BIN or
RECALL_CLI_BIN to point at a binary somewhere unusual, or set
[llm.cli] command.
Output shapes differ, so output_mode is part of the config. claude
prints prose (raw), gemini and grok print one JSON document
(single-json plus result_json_path), and codex --json prints one JSON
event per line (ndjson), where the answer is the last event matching
type=item.completed and item.type=agent_message. A future CLI with any of
those shapes is a config change:
[]
= "ndjson"
= ["type=item.completed", "item.type=agent_message"]
= "item.text"
Two codex-specific traps are handled by the preset, and matter if you write
your own [llm.cli] for it: --skip-git-repo-check is required, because
codex refuses to run outside a trusted git directory and a memory directory
usually is not one; and codex's -p is --profile, not the prompt — unlike
claude, gemini and grok — so the prompt goes in on stdin.
Caveats, stated plainly.
- gemini's result field is unverified. The flags were checked against
gemini0.27.3, but no authenticated call was made, so the preset triesresponse, thenresult, then falls back to raw stdout. If your Gemini wraps the answer in something else, set[llm.cli] result_json_path, or clearoutput_format_flagto take plain text instead. - In the background daemon, only file-based CLI auth works. The daemon is
started with an allowlisted environment (
PATH,HOME, …) that excludes API keys, so a CLI that authenticates through$HOMEkeeps working there while one that needsSOME_API_KEYexported behaves like the API providers and disables itself. See Background extraction.
Configuration
Optional .recall-echo.toml in the memory directory:
[]
= 5 # Rolling window size (1-50, default 5)
[]
= "claude-code" # See "LLM providers" above for the full list
= "" # Model name (provider default if empty)
= "" # Custom API base URL (HTTP providers only)
[] # Spawned-CLI providers only; every key is optional
= "" # Override the preset's binary
= "raw" # "raw" | "single-json" | "ndjson"
= "" # Where the answer sits in JSON output ("" = stdout)
= 300 # Per-call limit (0 waits forever)
[]
= "/path/to/journal" # Directory containing pipeline documents
= true # Auto-sync pipeline docs to graph on archive
[]
= "embedded" # Storage backend: "embedded" (default) or "server"
= "ws://localhost:8787" # SurrealDB server URL (server mode only)
[]
= 0.92 # At or above this cosine similarity, the same entity — no model call
= 0.82 # Below this, a new entity — no model call
= 3 # Existing entities compared per candidate
[]
= "" # Daemon socket override (default: XDG runtime dir)
= 3600 # Shut the daemon down after this much inactivity (0 = never)
[]
= true # Let the daemon extract entities when the machine is quiet
= 120 # Quiet period before a background batch starts
= 3 # Archives per batch (the next batch is one quiet period later)
| Section | Key | Default | Description |
|---|---|---|---|
ephemeral |
max_entries |
5 |
Rolling window size for session summaries (1-50) |
llm |
provider |
anthropic |
LLM backend — see LLM providers |
llm |
model |
provider default | Model name |
llm |
api_base |
provider default | Custom API base URL (HTTP providers) |
llm.cli |
preset |
from provider |
Calling convention: claude-code, gemini, grok, codex, custom |
llm.cli |
command |
preset binary | Binary name or path to spawn |
llm.cli |
args |
preset | Fixed arguments before the generated flags |
llm.cli |
prompt_delivery |
preset | stdin, flag or arg |
llm.cli |
prompt_flag |
preset | Flag carrying the prompt when delivery is flag |
llm.cli |
model_flag |
preset | Flag selecting the model (omitted when empty) |
llm.cli |
output_format_flag / output_format_value |
preset | Output-format flag and its value (empty value passes the flag alone) |
llm.cli |
output_mode |
preset | Stdout shape: raw, single-json or ndjson |
llm.cli |
ndjson_match |
preset | path=value predicates selecting the answer's line in ndjson mode |
llm.cli |
system_prompt_flag |
preset | Flag for the system prompt; empty prepends it to the message |
llm.cli |
result_json_path |
preset | Dotted path(s) to the answer in JSON output; empty = raw stdout |
llm.cli |
extra_args |
preset | Arguments appended after the generated flags |
llm.cli |
timeout_secs |
300 |
Per-call wall-clock limit (0 waits forever) |
pipeline |
docs_dir |
— | Path to pipeline documents (LEARNING.md, THOUGHTS.md, etc.) |
pipeline |
auto_sync |
false |
Sync pipeline documents to the knowledge graph on archive |
graph |
mode |
embedded |
Storage backend: embedded (single-process SurrealKV) or server (shared SurrealDB) |
graph |
url |
ws://localhost:8787 |
SurrealDB server URL, server mode only |
graph.dedup |
certain_similarity |
0.92 |
Cosine similarity at or above which a candidate is the same entity, resolved without a model call |
graph.dedup |
review_similarity |
0.82 |
Cosine similarity below which a candidate is a new entity, created without a model call |
graph.dedup |
max_candidates |
3 |
How many existing entities dedup compares a candidate against |
serve |
socket_path |
XDG runtime dir | Daemon socket path override |
serve |
idle_timeout_secs |
3600 |
Daemon idle shutdown timeout (0 disables) |
extraction |
background_enabled |
true |
Extract entities in the daemon once it has been quiet |
extraction |
idle_after_secs |
120 |
Seconds without a request before a background batch may start (0 = as soon as no connection is open) |
extraction |
batch_size |
3 |
Archives per background batch |
All settings have sensible defaults. Missing file or invalid values fall back silently.
Concurrency
The default embedded backend (SurrealKV) takes a process-exclusive file
lock on the graph store: one process may hold it at a time. recall-echo
resolves that with a small daemon rather than with locking rules you have to
remember.
How it works. The first graph command or hook that needs the store starts
a daemon in the background (recall-echo serve) and talks to it over a unix
socket in $XDG_RUNTIME_DIR/recall-echo/. Every later command — from any
number of concurrent sessions — goes through that same daemon, so concurrent
searches, queries and ingests all succeed. The daemon owns the store and the
embedding model, which also removes the per-command ONNX model reload. After
[serve] idle_timeout_secs of inactivity (default one hour) it exits.
- One daemon per memory directory: separate graphs never share a process.
- Crash-only: the daemon keeps no state outside the database. Kill it at any moment; the next command cleans up the dead socket and starts a new one.
- No silent fallback: if the daemon cannot start, the command fails with a
named error explaining what went wrong, including the tail of
<memory_dir>/graph/daemon.log. - Admin commands take the store exclusively.
graph init,gc,extract,ingest-all,pipeline status/flow/stale,vigil-syncanddecay-reporttake an admin lock beside the socket, stop the daemon and run in-process. A command that arrives while that lock is held waits for it instead of starting a daemon, and the lock is released only after the store is closed — so there is exactly one owner of the store at every instant. - Owner-only, both ends. The socket and its directory are
0700/0600and both ends check the peer's uid, so no other local user can read what you ingest or answer your queries. A[serve] socket_pathyou configure must already exist as a directory only you can write into: recall-echo validates it, but never creates or chmods a directory it did not derive itself. - Inspect it with
graph daemon status; stop it withgraph daemon stop.
External SurrealDB (advanced). For deployments that already run a
SurrealDB server (benchmark rigs, shared entity hosts), set
[graph] mode = "server" and point url at it. Commands then connect
directly and no daemon is involved. The backend is chosen at runtime — no
rebuild needed.
If the embedded store is locked by a foreign process, the daemon retries with
backoff and then fails with a named store locked error — never a raw LOCK
panic. Flat-file layers (MEMORY.md, EPHEMERAL.md, archives) are unaffected by
any of this.
Background extraction
Episodes arrive mechanically: the SessionEnd hook ingests every conversation
without anyone remembering to. Turning those episodes into entities,
relationships, confidence and provenance used to be a command a human had to
type — so semantic search and Bayesian confidence, the features this project is
actually about, stayed empty for anyone who did not read the docs closely.
The daemon does that pass itself. It already owns the store and already knows
when nobody is using it, so once the memory directory has been quiet for
[extraction] idle_after_secs (default two minutes) it takes batch_size
un-extracted archives (default three) and runs exactly what
recall-echo graph extract runs. Then it goes back to waiting. A backlog
drains one batch per quiet period.
- A client request always wins. Nothing is locked for the length of a batch: the worker shares the store like any other task, and it stops at the next archive boundary as soon as a connection is open.
- Interruption is safe. The
extractedflag flips per archive, after that archive succeeded — never per batch. A daemon killed mid-batch simply leaves work to do. - Idle shutdown waits for the batch, not for the backlog. The daemon will
not exit with an archive in flight, and a finished batch restarts the idle
clock — so a daemon working through a backlog stays up, and one with nothing
left to do exits after
[serve] idle_timeout_secsas always. - It gives up rather than loops. An archive that fails twice is written to
graph/extraction-quarantine.txtand skipped; three failures in a row disable the worker until the daemon restarts. There is no retry storm and no runaway bill. - It says what it did. Every batch logs its count and duration to
<memory_dir>/graph/daemon.log, andgraph daemon statusreports whether the worker is on, what it has extracted, and — when it is off — why.
What it costs. Extraction calls a model, so this is real money for
API-key providers. Two things bound it. First, the daemon is started with an
allowlisted environment that deliberately excludes API keys, so an
auto-started daemon can only ever use a CLI provider that authenticates
through $HOME — claude-code, and any other agent CLI with file-based auth —
which bills nothing beyond the subscription you already have; with
provider = "anthropic" the worker finds no key, disables itself, and says so
once in the daemon log. An API
provider reaches the daemon only if you run recall-echo serve --foreground
with the key exported — an explicit act. Second, batch_size bounds any single
burst. Set background_enabled = false to turn the pass off entirely.
Not in server mode. With [graph] mode = "server" clients bypass the
daemon completely, so a daemon's "quiet" measures a socket nobody connects to
rather than anything about you, and other processes may be writing to the same
store. Background extraction stays off there; graph extract is the way.
Contributing
See CONTRIBUTING.md for branch naming, commit conventions, and workflow.
License
MPL-2.0 — file-level copyleft. You may use recall-echo inside a closed-source product without opening your own code; modifications to recall-echo's own files must be published under the MPL.
Versions up to and including v3.13.0 were released under AGPL-3.0 and remain available under those terms.
Dependency note: the graph store uses SurrealDB 3.x, under the Business Source License 1.1 — source-available rather than OSI open source, converting to Apache-2.0 on 2030-01-01. Its use grant covers embedding the engine (what recall-echo does); it restricts offering SurrealDB itself as a database service to third parties.