Skip to main content

galet-memory

Resource metadata redesign: see the v3 schema review and migration plan. Semantic metadata v3 is opt-in for file/inline resources; existing configurations continue using v2. See the runtime and integration guide.

For application, environment-variable and shared-file configuration, see Configuring galet-memory. The public build_memory() entry point and galet-memory-configured sample use that same JSON contract.

Provider-neutral memory abstractions and reusable implementations for agent applications.

Galet-memory owns the neutral episodic, semantic, procedural, and working-memory contracts, plus reusable persistence and retrieval. An application decides when to record, recall, or archive; galet-memory determines how those operations work. It can run standalone without a host application. Galet supplies live model and embedding providers; offline samples use local fixtures.

The package includes concrete digest generation, non-destructive archive and reset boundaries, and semantic ingestion and recall. The older episodic CLI still accepts caller-supplied digest text as a separate example. See the galet architecture notes for the intended boundaries between the galet packages.

The CoALA memory model

Galet-memory implements the memory architecture described in the CoALA framework (Sumers, Yang, Narasimhan & Yao, Cognitive Architectures for Language Agents). CoALA describes a language agent as a decision cycle over four memory types, and this package provides a store for each of them:

CoALA memory Galet-memory store Role in the decision cycle
Working SqliteWorkingMemory Scratch state for the current run/task, with optimistic versioning and TTL expiry
Episodic SqliteEpisodicMemory / JSONL (SessionStore, EventStore, DigestStore) The agent's stream of experiences: sessions, events, exchanges, and their digests
Semantic VectorSemanticMemory + SemanticIngestionService Knowledge derived from those experiences, queried by meaning
Procedural FileProceduralMemory / ContextProceduralMemory Skills and rules for acting: contexts, imports, and scope precedence

CoALA assigns the decision cycle — when to record, recall, or archive — to the agent's policy, not the memory system. That is exactly the boundary this package keeps: the host application decides when an operation happens; galet-memory defines how it works and stays neutral about the agent around it. Companion packages handle the other CoALA layers: galet-prompt-builder turns recalled memory into bounded prompts, and galet supplies the model/embedding providers the memory stores plug into.

Install and test from a checkout

For the configuration features on develop, run these commands from the repository directory. Use Python 3.11 or later:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[test,vec]"
python -m pytest

If your virtual environment already exists, activate it and run the last two commands. Repeat the installation command after pulling changes that add or change dependencies. It installs the declared requirements, including jsonschema, and registers the command-line samples. A Git pull alone does not update the environment's dependencies.

For a released version from PyPI:

python -m pip install "galet-memory[vec]"

The new configuration commands require a checkout containing this change until it is included in a published release. Installing without [vec] omits the sqlite-vec dependency needed by semantic storage.

Configuration and running

A self-contained, copy-and-edit template is available at galet_memory_sample.json. It includes roots, an embedding profile, and settings for all four memory components in one file. Replace the placeholder absolute root directories and create the data and digests directories before assembling it. The template selects a live OpenAI profile; Galet supplies its credentials. The initialized offline example below creates its own directories and sample note.

Galet-memory is a library with command-line examples. The configuration sample assembles memory, performs ingestion and recall, and exits; it is not a background server.

Start with an offline example that needs no credentials. Choose a new directory; initialization refuses to overwrite an existing directory:

galet-memory-configured --init-sample "$HOME/galet-memory-demo"
export GALET_MEMORY_CONFIG_PATH="$HOME/galet-memory-demo/memory.json"
galet-memory-configured --offline

Initialization creates the directories, a Markdown note, and three JSON files:

File Purpose
memory.json Selects memory components, stores, embedding profile and keyword settings
roots.json Maps external-root names to absolute directories
embedding-profiles.json Defines provider, model and dimensions for each named profile

The sample uses the same JSON contract and public loader as applications and integration tests. It ingests the sample note and reports vector and keyword matches. Re-running with the same configuration reuses the stores and reports the unchanged source. Its fake provider has its own sample/offline-hash-v1 profile and semantic-offline.sqlite database.

You can select the configuration file explicitly instead:

galet-memory-configured --config "$HOME/galet-memory-demo/memory.json" --offline

Selection is explicit configuration object > explicit file > the file named by GALET_MEMORY_CONFIG_PATH. Sources are not silently merged. Relative JSON file references resolve from the containing config file, not the shell's current directory. New database files require create: true and an existing parent directory. Existing databases are validated, not automatically migrated.

Use a live embedding provider

In the generated memory.json, change memory.semantic.embedding_profile to semantic_default and change memory.semantic.database.relative_path to semantic-live.sqlite. The shared profile file already defines OpenAI's text-embedding-3-small with 1536 dimensions. Then configure Galet credentials and run without --offline:

export GALET_CREDENTIAL_PATH=/home/junwin/credential
galet-memory-configured --config "$HOME/galet-memory-demo/memory.json"

This makes paid embedding requests. Use a separate database for a different embedding profile; offline and real vectors are incompatible.

Environment variable Meaning
GALET_MEMORY_CONFIG_PATH Path to the memory JSON configuration file
GALET_CREDENTIAL_PATH Galet's existing directory of provider credential files

Galet also supports its existing provider-key environment variables and explicit Settings. No GALET_CONFIG_PATH general config loader is introduced here. See Configuring galet-memory for complete JSON examples, Galet integration and resource-access behavior.

GALET_MEMORY_CONFIG_PATH is used by load_memory_config(), build_memory(), and galet-memory-configured. The older commands below keep their explicit arguments and do not read this JSON configuration.

Quick sample

A no-credential working-memory example — start a run, store a value with optimistic versioning, finish the run (uses a temporary database):

galet-memory-working-road-test

Further runnable examples (episodic CLI, procedural recall, semantic and digest road tests, and a keyword extraction check) are described in the sections below.

Dependency rule

Applications may depend on galet-memory; galet-memory must never depend on an application.

In particular, this package never imports host-application code. The host owns application-specific configuration merging, agent lookup, request context and access policy. Galet-memory owns validation of its JSON contract and assembly of its reusable components through build_memory(); applications may also supply dependencies explicitly.

The episodic interfaces are explicit session/event contracts; see docs/design/episodic-interfaces.md.

Storage ports

Reusable memory implementations depend on narrow, structurally typed ports:

  • EmbeddingProvider turns text into vectors. GaletEmbeddingProvider adapts Galet's EmbeddingApi.
  • EmbeddingIndex performs namespace-scoped similarity queries without exposing an application's storage records.
  • TextLoader loads bounded text from paths already authorized by the host.
  • KeywordExtractor derives keywords from text. SimpleKeywordExtractor is a dependency-free default; applications may supply their own (for example spaCy/NLTK-based) through the same port.
  • ContextRepository returns neutral context and skill snapshots.

Host applications adapt their storage and configuration to these ports. The package includes VectorSemanticMemory, EmbeddingDigestRecall, ContextProceduralMemory, and a basic FileTextLoader.

Working memory lifetime

SqliteWorkingMemory holds temporary JSON state under an account and run ID, optionally scoped to a task. The application starts a run with a fixed TTL and passes that identity to agents that should share state. Writes require an expected version; a stale update fails instead of overwriting another agent's work. Missing keys return None, while a finished or expired run raises WorkingRunUnavailable. finish_run removes values immediately; expiry is enforced on reads and writes even before purge_expired removes old rows.

from galet_memory import SqliteWorkingMemory

with SqliteWorkingMemory("/path/to/working.sqlite") as working:
    working.start_run("alice", "run-uuid", ttl_seconds=3600)
    plan = working.put("alice", "run-uuid", "plan", {"next": "review"},
                       expected_version=0)
    working.put("alice", "run-uuid", "plan", {"next": "report"},
                expected_version=plan.version)
    working.finish_run("alice", "run-uuid")

The database is a temporary-state store, separate from procedural contexts, skills, session events, and the host's durable task storage. Run identifiers and access to them remain the host application's responsibility. The default maximum value is 64 KiB. Run a disposable handoff and expiry example:

galet-memory-working-road-test

Procedural memory locations

FileProceduralMemory reads and writes Markdown contexts and skills under a root chosen by the application. The default ProceduralLayout has global, account, and account/project locations; same-named contexts contribute text in that order. Imported skills resolve to the most specific definition. Frontmatter supports imports, tag, mandatory_tools, and search_namespaces. The result includes source paths and scopes.

from galet_memory import FileProceduralMemory, ProceduralMemoryRequest

memory = FileProceduralMemory("/srv/agent-data")
result = memory.recall(ProceduralMemoryRequest(
    account_name="alice", context_name="shop", project_name="boutique"
))

Recall is read-only by default. Applications explicitly call memory.repository.save_context(...) or save_skill(...) to create files. For Lucy's existing contexts/<account>/*.md and skills/<account>/*.md layout, pass ProceduralLayout.lucy() and point root at the directory containing contexts and skills.

Contexts default to context_resolution="merge" for compatibility. Pass context_resolution="most_specific" to FileProceduralMemory (or FileContextRepository) to select only the most specific existing context. This replaces the entire definition, including imports, tools, tag, and search namespaces; an empty account file still overrides the global file. Skills always select the most specific existing definition independently of this policy.

Custom ProceduralLayout paths can preserve an existing directory structure:

layout = ProceduralLayout(
    global_contexts="contexts", account_contexts="contexts/{account}",
    project_contexts=None,
    global_skills="skills", account_skills="skills/{account}",
    project_skills=None,
)
memory = FileProceduralMemory(root, layout, context_resolution="most_specific")

repository.list_resolved_context_names(account) lists visible names across scopes without duplicates. read_effective_context reads the most specific raw file; resolve/recall applies the configured context policy. Existing scope-specific read/list methods keep their original behavior. To edit an inherited context without modifying a shared file, use update_context(..., scope="account", inherit_existing=True); its body and frontmatter are copied from a less-specific definition before applying changes. Writes still target the explicit scope. Recall and listing do not create files.

Run a disposable example, or inspect existing files without changing them:

galet-memory-procedural
galet-memory-procedural --root /path/to/storage/data --layout lucy \
  --account alice --context shop

The root and layout are supplied when composing the application; request fields select the account, context, and optional project. Galet-memory owns path validation, Markdown parsing, imports, and scope precedence.

Request-scoped embedding reuse

Semantic document recall and episodic digest recall commonly embed the same query. Wrap their shared provider once and open a cache scope around the application request:

from galet_memory import CachingEmbeddingProvider

embeddings = CachingEmbeddingProvider(base_embeddings)
semantic_memory = VectorSemanticMemory(embeddings=embeddings, ...)
digest_recall = EmbeddingDigestRecall(embeddings=embeddings, ...)

with embeddings.request_scope() as cache:
    digest_recall(digest_search_request)
    semantic_memory.recall(semantic_request)

print(cache.info())

The cache key contains the exact embedding model and the exact ordered input texts. A model change therefore cannot reuse vectors from the previous model. Calls outside a request scope pass through uncached, concurrent contexts are isolated, and failed provider calls are never stored.

SqliteVecEmbeddingIndex reads the existing 1536-dimension sqlite-vec schema. It defaults to vec_embeddings_v2 and joins results to embedding_metadata; the original vec_embeddings table can be selected explicitly for legacy reads. Internal tables generated by vec0 are never accessed directly.

Explicit episodic interfaces

SessionStore, EventStore, and DigestStore have named operations in episodic/session_interface.py, event_interface.py, and digest_interface.py. A session belongs to an account and can contain messages from several agents. An event records its role, actor, kind, complete content and correlation IDs. The caller does not select an agent to read or manage a conversation.

from galet_memory import SqliteEpisodicMemory, NewEvent

with SqliteEpisodicMemory("/path/to/new-chat.sqlite") as memory:
    session = memory.create_session(account_name="junwin")
    memory.append_event(
        account_name="junwin", session_id=session.session_id,
        event=NewEvent("user", "Hello", "junwin", correlation_ids=("request-uuid",)),
    )
    page = memory.get_recent_events(
        account_name="junwin", session_id=session.session_id, count=10,
        event_kinds=["user_message"],
    )
    for event in page.events:
        print(event.content, event.correlation_ids)

Filters apply before counts. Recent reads return the last N matching active records in append order; reports and tool calls do not consume a user-message limit. Period reads use timezone-aware [start, end) boundaries and return occurrence-time order with append sequence as the tie break. Event reads do not count tokens or truncate content. Galet-prompt-builder owns prompt budgets.

get_session reads metadata only. SessionChanges expresses typed updates; EMPTY_TAIL explicitly guards an empty event stream. Every operation is scoped to the account. get_exchange reads linked request/response/processing events; invalidate_exchange hides them idempotently and excludes affected digests. See event invalidation.

Both SQLite and JSONL implement the same public contracts. This is a breaking interface/storage change: use a fresh database or directory. Existing episodic schemas/layouts are rejected; there are no migrations, compatibility readers or combined recall API. SQLite storage describes the relational schema. JSONL uses a complete snapshot replaced atomically under a cross-process lock; it is intended for small stores.

Episodic memory road test

Inspect user messages and exchanges

Use a fresh path for the new episodic schema. The account is explicit for every command; session metadata does not require an agent. Actor identifies the producer of each appended event.

galet-memory-episodic --db /tmp/galet-chat-new.sqlite --account demo create --session-id road-test
galet-memory-episodic --db /tmp/galet-chat-new.sqlite --account demo append road-test "Hello" --actor demo --correlation request-1
galet-memory-episodic --db /tmp/galet-chat-new.sqlite --account demo append road-test "Welcome" --role assistant --actor peace --correlation request-1
galet-memory-episodic --db /tmp/galet-chat-new.sqlite --account demo recent road-test --count 10 --kind user_message
galet-memory-episodic --db /tmp/galet-chat-new.sqlite --account demo exchange road-test request-1
galet-memory-episodic --db /tmp/galet-chat-new.sqlite --account demo invalidate road-test request-1

Other commands include list, show (session metadata), event, period, update, clear, delete and the supplied-text archive demonstration. Use --help on the CLI or a subcommand for its arguments. Period timestamps require a timezone, for example --start 2026-10-01T00:00:00Z --end 2026-10-02T00:00:00Z.

Check the digest lifecycle

Run the fixture-based archive, cumulative digest and reset checks with fresh temporary storage:

galet-memory-digest-road-test

To use a real model through Galet, supply --live --model gpt-4o-mini and, if needed, --credential-path /path/to/credentials. Live mode makes paid requests.

Semantic ingestion and recall

The examples in this section use the lower-level constructor/CLI APIs. The configured path above always supplies keyword extraction and embedding-profile checks; direct constructor callers choose those dependencies explicitly.

SemanticIngestionService owns add, refresh, unchanged-content detection, and targeted removal of text or caller-authorized UTF-8 files. It uses the same EmbeddingProvider and EmbeddingIndex ports as VectorSemanticMemory. Sources are identified within an account and namespace; files are recalled through FileTextLoader, while directly ingested text is stored for recall in record metadata. An oversized source is rejected rather than silently truncated during embedding. The current SQLite vec schema requires 1536 dimensions, so use a matching embedding model.

When the service is constructed with a keyword_extractor, keywords are derived alongside the embedding vector at every ingest and stored with the record. Existing records are unaffected: without an extractor no keywords are written, and records ingested before keywords exist refresh once with keywords on their next re-ingest.

ingest_folder walks a caller-authorized folder and ingests every matching file with the same stable identity per resolved path, so unchanged files skip the embedding call on re-runs:

results = ingestion.ingest_folder(
    account_name="alice", namespace="docs", folder="/srv/agent-data/notes",
    suffixes=(".md", ".txt"), recursive=True,
)

By default an unreadable or oversized file stops the walk; pass skip_errors=True to continue past individual failures.

Run a complete add → recall → edit → refresh → recall → delete sequence in a temporary SQLite vec database without credentials:

pip install -e '.[vec]'
galet-memory-semantic-road-test

To make real embedding calls through Galet, set OPENAI_API_KEY or supply a credential directory, then run:

galet-memory-semantic-road-test --live --model text-embedding-3-small \
  --credential-path /path/to/credentials

Live mode makes paid embedding requests. Both modes check account isolation, unchanged source handling, refresh, and deletion.

Install the package, set OPENAI_API_KEY (or use Galet's GALET_CREDENTIAL_PATH), and point the CLI at a copy or test embedding database.

Add a sample:

galet-memory-embeddings \
  --db /path/to/embeddings.sqlite \
  --account junwin \
  --namespace demo \
  add "The allotment has runner beans and three apple trees."

Query it:

galet-memory-embeddings \
  --db /path/to/embeddings.sqlite \
  --account junwin \
  --namespace demo \
  query "What fruit trees are in the allotment?"

Query it with path to credentials:

galet-memory-embeddings \
  --credential-path /path/to/credentials \
  --db /path/to/embeddings.sqlite \
  --account junwin \
  --namespace demo \
  query "What fruit trees are in the allotment?"

Use --extension when vec0 is not installed at /usr/local/lib/sqlite-vec/vec0.so. The current schema requires 1536-dimension vectors, so the default model is text-embedding-3-small. Run against a copy of a production database when experimenting.

Keywords and fallback recall

Keywords give semantic memory a degradation path when an embedding service is unavailable, and a source of derived terms for each stored record.

Ingestion

SemanticIngestionService(keyword_extractor=SimpleKeywordExtractor()) derives keywords from every ingested text or file alongside the embedding vector. The KeywordExtractor port is structural; SimpleKeywordExtractor is a dependency-free default (token frequency with a small built-in stopword list, extendable via stopwords=(...)). Applications can supply their own extractor — for example a spaCy/NLTK-based one — without changing galet-memory.

Keywords are stored in the keywords column of embedding_metadata (a JSON array). The lower-level SqliteVecEmbeddingIndex constructor, with its default initialize_schema=True, can add a missing keywords column via ALTER TABLE. The configured build_memory() path opens existing stores without schema initialization and rejects a missing keywords column; it requires an explicit migration. Records whose keywords exist only inside source_metadata (from earlier ingestion) can still be matched by the lower-level keyword query.

Backfilling existing records

galet-memory-embeddings backfill-keywords derives keywords for records that were ingested before keyword extraction existed — without re-embedding:

galet-memory-embeddings \
  --db /path/to/embeddings.sqlite \
  --account junwin \
  --namespace demo \
  backfill-keywords

It reads each record's stored inline text or recorded path, extracts keywords with SimpleKeywordExtractor, and writes them with a metadata-only update: no vector writes and no embedding API calls. Records with no reachable text (for example a moved or deleted file) are skipped and counted. Run it against a copy of a production database first.

Fallback recall

VectorSemanticMemory(keyword_extractor=...) embeds the query and queries by vector as usual. If the embedding provider fails (network, quota, outage) and a keyword extractor is configured, recall degrades to keyword matching instead of raising:

memory = VectorSemanticMemory(
    embeddings=embeddings, index=index,
    text_loader=FileTextLoader(),
    keyword_extractor=SimpleKeywordExtractor(),
)
result = memory.recall(SemanticMemoryRequest(
    account_name="alice", query="apple trees", namespaces=["notes"],
))
if result.metadata["backend"] == "keyword":
    # embedding service was unavailable; scores are keyword-overlap fractions
    ...

Keyword results are marked metadata={"backend": "keyword", "reason": "embedding_provider_unavailable", "query_keywords": [...]} so callers can distinguish them from vector results (backend="vector"). Keyword scores are the fraction of query keywords present in a record's stored keywords; they are not comparable with embedding similarity scores, and score_threshold applies to them unchanged. Recall with use_embeddings=False also routes to keyword recall when an extractor is configured. Without a configured extractor the previous behavior is preserved: provider failures propagate and use_embeddings=False returns an empty result.

Run a disposable check of the default extractor:

galet-memory-keywords-road-test

Config-driven integration runner and semantic metadata v3

galet-memory-integration runs ingestion/recall and enabled memory-component checks using your configuration, with all writable storage redirected to a new test workspace. It defaults to offline embeddings; --live explicitly selects paid embeddings from the configured provider. V3 file/inline metadata support is opt-in; omitted metadata_version retains v2. See the v3 runtime and integration guide for commands, migration verification, and remaining digest/event/Lucy integration work.

Configuring existing procedural layouts

memory.procedural.layout accepts the six ProceduralLayout fields: global_contexts, account_contexts, project_contexts, global_skills, account_skills and project_skills. Values are path templates or null to disable that scope. Omitted fields retain galet-memory defaults. For an existing contexts/<account> and skills/<account> tree, set the corresponding templates to contexts/{account} and skills/{account} explicitly. This configuration is passed by build_memory to the same procedural backend used by direct callers.

Metadata

Release files for galet-memory 0.1.29

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for galet-memory 0.1.29
File Size Uploaded Provenance Actions
galet_memory-0.1.29.tar.gz 126.5 kB Without provenance Details for galet_memory-0.1.29.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for galet-memory 0.1.29
File Interpreter ABI Platform Actions
galet_memory-0.1.29-py3-none-any.whl Python 3 none any Details for galet_memory-0.1.29-py3-none-any.whl

Total release size: 234.2 kB

Release files / galet_memory-0.1.29.tar.gz

Download URL galet_memory-0.1.29.tar.gz
Size 126.5 kB
Tags Source
SHA-256 checksum
How to use checksums
16bc352d8cac94c6426b2c411d3f211e24e5a788782cf3a5c99380a8c5b3a776
BLAKE2b-256 checksum
How to use checksums
adefa892956e8e2ead543867150f9c739d2b0eb2417f5f28a2dbb3be691b6754
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via uv/0.13.5 {"installer":{"name":"uv","version":"0.13.5","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Linux Mint","version":"22.3","id":"zena","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":null}

Release files / galet_memory-0.1.29-py3-none-any.whl

Download URL galet_memory-0.1.29-py3-none-any.whl
Size 107.7 kB
Tags Python 3
SHA-256 checksum
How to use checksums
5e69eefc9ffd72ad980bbd4970ff279bfa19c3dc358630805d2f9baf3c7b5918
BLAKE2b-256 checksum
How to use checksums
2e37a1303e2b8d94a5e207eb2cf98f3d3df8e938bbd6fd99e019b6fb3fd2ed88
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via uv/0.13.5 {"installer":{"name":"uv","version":"0.13.5","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Linux Mint","version":"22.3","id":"zena","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":null}

Security for galet-memory 0.1.29

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Release history Release notifications | RSS feed

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page