Skip to main content
Glama

tether

A shared memory layer for personal agents, across devices. tether is an MCP server backed by a local SQLite file. Any MCP-compatible agent can remember, recall, link, and forget durable notes — facts about you, your projects, your preferences — so context follows you instead of dying with each session.

It runs local-only with zero configuration. Point it at a hosted libSQL/Turso primary and the same file becomes an embedded replica that syncs your memory across every device in near-real-time.

Why

The near future is personal agents living across many devices — laptop, desktop, phone. For that to feel like one assistant rather than several amnesiac ones, memory has to be a substrate that follows you: readable and writable from every device and from any agent, not siloed inside a single tool.

tether is that substrate. It is deliberately a convenience layer — it makes an agent more useful when present, and never breaks the agent's work when degraded.

Related MCP server: CP Memory

Status

v0.6.2. The core (four memory verbs + boot index + FTS5) shipped in v0.1; since then recall has grown a semantic arm, consolidation, an associative usage graph, a self-organizing store, and opt-in crystallization — each additive and each degrading cleanly to plain keyword recall. Every feature below is implemented.

Design and rationale start at docs/superpowers/specs/2026-07-03-tether-design.md; the associative core, seed-dominant recall, self-organizing store (Tier B1), and crystallization (Tier B2) each have their own design doc under docs/superpowers/specs/, with matching plans in docs/superpowers/plans/.

Design at a glance

  • Four memory verbs: remember · recall · link · forget. (Enabling crystallization adds one reflection-control tool, dismiss_cluster — not a memory operation.)

  • Upsert on write so the store doesn't rot into near-duplicates.

  • Rich recall (id, type, title, body, tags, updated_at, plus a via receipt saying why each hit surfaced) so an agent can judge staleness and cite what it updates. body is a query-centered excerpt, not the whole memory — see Excerpts.

  • An auto-loaded boot index — a compact one-line-per-memory list surfaced to the agent each session, so memory helps even when the agent doesn't think to search.

  • Local-first, sync optional — the local path is untouched when no backend is configured; degradation never throws.

  • Hybrid search, associative on top — FTS5 keyword hits and local static embeddings are fused, then a usage graph (explicit links, learned co-recall, semantic neighbours) pulls in connected memories. Every layer is additive and degrades to plain keyword recall.

  • Safe under parallel tool calls — an agent that fires several remember/recall calls at once gets atomic, correctly-reported results; see Performance and durability.

Install

Requires Python ≥3.10 on Linux, macOS, or Windows.

Register it with Claude Code — with uv:

claude mcp add tether -- uvx tether-memory

…or install it first:

pip install tether-memory
claude mcp add tether -- tether-memory

(The package is named tether-memory on PyPI — tether was already reserved as a common brand name. tether in claude mcp add tether -- ... is just the label Claude Code uses to refer to this server; it doesn't need to match the installed command.)

Or add it as a Claude Code plugin (this repo doubles as a one-plugin marketplace; the plugin runs uvx --from 'tether-memory[semantic]', so semantic recall comes along):

/plugin marketplace add sidyellur/tether
/plugin install tether@tether

It is also listed in the MCP Registry as mcp-name: io.github.sidyellur/tether.

By default memory lives in a local SQLite file at ~/.local/share/tether/memory.db — on Windows, %LOCALAPPDATA%\tether\memory.db (override either with TETHER_DB, or set XDG_DATA_HOME, which is honored on every platform). No accounts, no network — this is the whole tool for a single machine.

Project awareness

tether knows which project it is serving: Claude Code sets CLAUDE_PROJECT_DIR in every MCP server's environment, and tether takes the directory's name as the project (override or disable with TETHER_PROJECT). That drives three things, none of which need configuration:

  • The boot index leads with this project. The auto-loaded index opens with a # This project (<name>) section, then # Everything else, so the agent starts a session already looking at the decisions and gotchas for the repo it is in rather than whatever you touched last, anywhere.

  • Work memories are tagged automatically. project, feedback and reference memories get a proj:<name> tag unless the agent passes a proj: tag itself. user memories are about you, not the work, and stay global. The tag is an ordinary tag: recall(tags="proj:<name>") lists a project's memories deterministically.

  • Recall prefers this project on a near-tie. A hit tagged with the current project ranks a few places ahead of an equally-good hit from another project; it never outranks a clearly better match, and untagged memories are neither boosted nor penalized.

Var

Default

Effect

TETHER_PROJECT

basename of CLAUDE_PROJECT_DIR

name the project explicitly; off disables project awareness

Nothing falls back to the working directory: outside Claude Code (or with TETHER_PROJECT=off) tether behaves exactly as before.

Sync across devices (optional)

Point tether at a Turso / libSQL database and the local file becomes an embedded replica — local-speed reads, writes that propagate to your other devices. Install the extra and set two env vars:

pip install 'tether-memory[sync]'
export TETHER_SYNC_URL='libsql://<your-db>.turso.io'
export TETHER_SYNC_TOKEN='<your-auth-token>'

If the backend is unreachable, tether logs sync offline and keeps working against the local file. Sync is a single-primary design: every write goes to the hosted primary and replicas only pull, so a device that was merely offline catches up (including tombstones from forget) on its next pull and can never resurrect a memory another device forgot. The one gap: writes made while degraded land only in that device's local file and are not merged back later (#99 tracks reconciliation).

Writes push immediately. Reads also pull, debounced to at most once every TETHER_SYNC_READ_INTERVAL seconds (default 30) — so a device that only asks things still sees what your other devices wrote, instead of staying frozen at its own startup state until it happens to write something. The read-path pull is bounded much more tightly than the write-path one: if the backend is slow, the recall returns local data and the pull lands for the next read rather than making you wait.

Var

Default

Effect

TETHER_SYNC_READ_INTERVAL

30

seconds between read-path pulls; 0 = only sync on writes

TETHER_DEVICE_ID

hostname

the device id recorded on each memory (and the default TETHER_AUTHOR)

One thing to know about replicas: libSQL forwards every write to the hosted primary, and with the associative graph on (the default) recall writes too — it records what was recalled together so memories can wire up over time. On a replica that makes each recall a few network round-trips on top of the local search. If that matters more to you than learned associations, TETHER_ASSOC=0 makes recall read-only again.

The keyword arm is SQLite FTS5 over title, body and tags, ranked by bm25. Ask in plain language: a memory that contains some of the query's words is a hit, and one that contains more of them ranks higher, so "how do we run the integration tests?" finds the note that says "pytest runs the tests" (common function words are ignored). The index stems English words, so tests matches test and deciding matches decided. Stemming is English-only; turn it off for a store in another language and tether rebuilds the index on the next start.

Var

Default

Effect

TETHER_FTS_STEMMING

on

set 0/false/off to index words exactly as written

Measured on the LoCoMo long-conversation benchmark (one memory per dialogue turn, 1,536 questions, "did recall return the turns that answer it" in the top 10), the keyword arm alone finds the evidence for 61% of questions, against 54% for a textbook BM25 over the same text. That harness ships in the repo — see Benchmarks.

Semantic search (optional)

By default recall is hybrid: keyword (FTS5) results are fused with semantic (vector) results, so a query finds relevant memories even when the exact words differ ("automobile" recalls a note about your "car"). Semantic recall runs a small static embedding model locally — no network, no API key, nothing to hang on. Install the extra:

pip install 'tether-memory[semantic]'

Without the extra (or with TETHER_SEMANTIC=0), tether runs keyword-only FTS5 — semantic is a pure add-on and never a requirement. The first run embeds existing memories once (a one-time backfill); after that it is incremental.

Environment:

Var

Default

Effect

TETHER_SEMANTIC

on

set 0/false/off to force keyword-only recall

TETHER_EMBEDDING_MODEL

minishlab/potion-base-8M

override the local static model

Consolidation (optional)

tether keeps a superseded fact rather than overwriting it: when a memory is replaced, the old one is marked no longer current (retained for history) and excluded from recall and the boot index. Recall also gently favors more recent facts. Two opt-in behaviors go further:

Var

Default

Effect

TETHER_CONSOLIDATE

off

on (1/true) merges a near-duplicate on write — supersedes the old fact instead of fragmenting the store (needs the [semantic] extra)

TETHER_DEDUP_THRESHOLD

0.92

cosine similarity required to treat two facts as duplicates

TETHER_DECAY_HALF_LIFE_DAYS

off

set a positive number to exponentially down-rank older facts in recall

TETHER_AUTHOR

device id

attribution recorded on each memory

Consolidation never deletes — forget soft-deletes (see Tools), and only the admin CLI's purge is permanent. All of this degrades to plain keyword recall when the semantic extra is absent.

Associative recall (optional)

recall doesn't just return keyword/semantic matches — it follows a usage graph to related memories, so asking about one thing surfaces its connected context. The graph's edges come from three local, deterministic sources — no LLM, no network:

  • semantic — nearest neighbours by embedding (needs the [semantic] extra),

  • explicit — the link() verb,

  • hebbian — memories you recall together get wired together over time.

Every hit carries a via receipt saying why it surfaced (a direct match, or the edge it came through), and two optional recall args tune it:

Arg / var

Default

Effect

budget (per call)

TETHER_RECALL_BUDGET

how far to follow associations; 0 = direct matches only

session (per call)

time-bucketed

group related recalls so they prime each other

TETHER_ASSOC

on

set 0/false/off for plain keyword+semantic recall

TETHER_RECALL_BUDGET

8

default association breadth

TETHER_PROTECT_HEAD

8

how many top direct hits are locked above associations

TETHER_SEED_FLOOR

0.35

minimum cosine similarity a semantic hit needs to seed an associative walk; below it a memory is only reachable through an edge. 0 disables the floor

Associative recall is seed-dominant: the top direct matches are locked in place, and associations only fill the slots below them — so turning association on never demotes a hit that keyword/semantic search already ranked highly.

With TETHER_ASSOC=0 (or budget=0, or an empty graph), recall behaves exactly as before — associative recall is purely additive and never breaks a lookup.

Self-organizing store (optional)

As a store grows, tether keeps it legible using the same usage graph:

  • Hub-curated boot-index. The auto-loaded memory index is capped once it passes TETHER_BOOT_INDEX_CAP (default 50) — the cap always applies, so a large store never gets an unbounded index. With a graph, above the cap it shows two labeled slices — load-bearing memories (highest behavioral degree: explicit links + learned co-recall, never mere similarity) and the most recent ones — so the index stays small and shows what actually matters. Without a graph (TETHER_ASSOC=0), it falls back to a plain most-recent-N list instead of the hub/recency split. Below the cap it's the full newest-first list either way.

  • Forgetting-by-disconnection (opt-in, TETHER_FORGET). A bounded sweep runs every TETHER_FORGET_INTERVAL writes and soft-archives memories that are both old (TETHER_FORGET_AGE_DAYS, default 90) and behaviorally isolated (no explicit/hebbian edge — semantic similarity doesn't count). Archived memories drop out of recall and the boot-index but are retained and reversible (it reuses the same mark-invalid machinery as consolidation; nothing is deleted). Safety rails: never runs without a live behavioral graph, below 2 × CAP memories, or more than TETHER_FORGET_MAX_PER_SWEEP (default 10) per sweep.

var

default

effect

TETHER_BOOT_INDEX_CAP

50

curate the boot-index above this size

TETHER_FORGET

off

enable the forgetting sweep

TETHER_FORGET_AGE_DAYS

90

minimum age to be eligible to fade

TETHER_FORGET_INTERVAL

20

writes between sweeps

TETHER_FORGET_MAX_PER_SWEEP

10

max archived per sweep

With TETHER_FORGET off (default) and a normal store size, recall and the boot-index behave exactly as before.

Crystallization (optional, off by default)

With TETHER_CRYSTALLIZE=1, tether reflects: it detects dense clusters of related memories and offers them for naming. Read tether://crystallization during a reflection pass (it is pull-only, never auto-loaded) to get candidate clusters; name a real principle with remember(..., crystallizes=[source_ids]) — which writes the principle and links it over its sources — or drop a candidate with dismiss_cluster(id_a, id_b). Clusters are seeded by explicit links + usage (semantic similarity fills out membership), so this finds "these belong together" structure, not mere topical similarity. tether finds the structure; your agent supplies the words.

A crystallized principle becomes a boot-index hub and is reachable from its sources in recall. Note: this makes "named" a third importance signal alongside "used" and "linked" — deliberate, since an agent judging something principle-worthy is a strong signal.

Tools

Tool

What it does

remember(type, title, body, tags?, links?, crystallizes?)

Save a memory; upserts on type+title so facts refine rather than duplicate. crystallizes=[ids] writes it as a principle over those sources (needs TETHER_CRYSTALLIZE)

recall(query?, type?, limit?, budget?, session?, tags?, id?, full?)

Hybrid keyword + semantic search, then follows the usage graph to related memories; returns id/type/title/body/tags/updated_at + a via receipt. body is a query-centered excerpt — see Excerpts — with id=N fetching one memory whole. tags is an exact-match filter (a memory must carry every listed tag); combine it with query, or omit query for a guaranteed-complete tag lookup

link(id_a, id_b)

Bidirectional link between two memories

forget(id)

Soft-delete a memory: marks it no longer current (excluded from recall/the boot index) via the same reversible valid_to machinery as consolidation, rather than deleting the row. See Export and permanent deletion for a real, permanent delete

dismiss_cluster(id_a, id_b)

Reflection control (crystallization): drop the candidate cluster nucleated by peak edge (id_a, id_b) so it isn't re-surfaced. Not a memory operation; only relevant with TETHER_CRYSTALLIZE

Plus three resources: the auto-loaded tether://memory-index (a compact one-line-per-memory index surfaced each session), the pull-only tether://status (runtime config: semantic/sync state, memory and edge counts, DB path — for debugging what's actually active), and, with TETHER_CRYSTALLIZE, the pull-only tether://crystallization (candidate clusters for a reflection pass).

Excerpts

recall returns a relevance-centered excerpt of each memory's body, not the whole thing — the window is centered on the stretch where the most of your query's words appear together (a word that is rare in that memory counting for more than one that is everywhere in it), so you see why the memory matched rather than just its opening lines. A hit that was cut also carries truncated: true and body_chars (the real length), and you fetch the one memory you actually want in full with recall(id=N).

This is the search-engine shape: the result list is an index of pointers with enough text to judge relevance, not a payload of documents. It matters because memories can be large — a single 44KB journal memory made unrelated queries cost ~57–67KB per call while the retrieval itself took under a millisecond. The response was fat, not the engine:

query

full bodies

excerpts

seed dominance

57.0KB

1.4KB

hebbian edges

66.9KB

2.0KB

cold start latency

66.9KB

2.0KB

A memory shorter than the excerpt width is returned whole and unmarked, exactly as before.

Var / arg

Default

Effect

TETHER_EXCERPT_CHARS

500

excerpt width; 0 returns full bodies

id (per call)

—

fetch just this memory, whole

full (per call)

false

full bodies for every hit — costs the whole payload; prefer id=

Performance and durability

tether is meant to be invisible in an agent's loop, so the hot paths are measured and kept flat as the store grows. Numbers below are from a local SQLite store with the semantic and associative layers on, single process:

memories

remember

recall (rare term)

recall (term in most memories)

500

1.6 ms

1.8 ms

3.7 ms

2,000

1.8 ms

2.5 ms

7.7 ms

8,000

3.4 ms

1.0 ms

19 ms

A few things that make this hold:

  • Writes don't scale with the store. The embedding matrix used for semantic search and neighbour wiring is kept in memory and patched row by row on every write, rather than re-read from SQLite. It is rebuilt only when vectors change wholesale (a model change, a backfill) or when another process has written to the file (a CLI purge, a second server, a sync pull) — SQLite's data_version counter catches that.

  • Parallel tool calls are serialized. MCP runs each tool call on its own thread, and agents issue calls in parallel. All Store operations take one lock, so a recall and a remember arriving together are each atomic: no interleaved transactions, no half-committed writes, and action is always right.

  • Commits don't fsync. Local connections run WAL with synchronous=NORMAL: still safe against corruption, but the last few transactions can be lost if the machine loses power before a checkpoint (an application crash loses nothing). Every remember and, with the graph on, every recall commits, so this is one disk sync saved per call.

  • Search stays cheap. Vector search is a single numpy matmul over the in-memory matrix — well under a millisecond at thousands of memories, which is why there is no vector-index extension to install. Keyword cost is FTS5's: proportional to how many memories match the query.

Costs to expect once: the first boot after installing the [semantic] extra (or changing the model) embeds every existing memory and wires its neighbours, which takes a second or two per few thousand memories. The boot index reads only the rows it prints; once the store is larger than the index cap it also works out which memories are most linked, which reads the link table once per session. Tag-only lookups go through the tag index.

Benchmarks

Two harnesses live in bench/, both runnable without an API key:

  • python -m bench.locomo — retrieval-only evaluation on LoCoMo, ten long multi-session conversations with ~1,500 questions each labelled with the dialogue turns that answer it. Every turn becomes a memory; the score is whether recall(question) returns those turns (recall@k, MRR), per condition: keyword only, keyword + semantic, and the full associative path, next to a textbook BM25 baseline. No LLM is involved, so the number measures the one thing a memory layer controls — did it hand the agent the right facts — and is not comparable to the LLM-judged "accuracy" figures memory vendors publish on the same dataset. The data (~1.5 MB) is downloaded on first use. Install the [semantic] extra to measure the semantic and associative conditions with the real model.

  • python -m bench.run — tether's own associative-recall evaluation: a controlled corpus of tasks whose members are used together, measuring what the usage graph adds over keyword + semantic search after simulated use. Needs the [semantic] extra.

Export and permanent deletion

forget never deletes data — it soft-deletes, like consolidation. Two admin operations, deliberately kept off the MCP tool surface so an agent can't trigger them, live in a small CLI instead:

tether export                    # dump all current memories to JSON (stdout)
tether export -o backup.json     # ...or to a file
tether import backup.json        # merge an export back into the store
tether restore <id>              # un-forget a soft-deleted memory
tether history <id>              # earlier versions of a memory that updates replaced
tether revert <id> <version>     # put one of those versions back
tether purge <id> --yes          # permanently delete a memory (bypasses forget)

import replays records through the normal write path, so it upserts on type+title like remember does — importing into a non-empty store merges rather than duplicating. Ids are not preserved (an id in the file may map to a different one here); links are remapped accordingly, and a link pointing outside the file is dropped rather than pointed at the wrong memory. The report tells you what happened: {"created", "updated", "skipped", "linked", "dropped_links"}.

restore clears valid_to, reversing a forget (or a consolidation, or a forgetting sweep). It refuses if a newer memory has since claimed the same type+title, naming the blocker rather than failing opaquely.

history lists what updates replaced. remember updates a memory in place when the title matches, but the previous title, body and tags are kept as a version (the newest 10 per memory), so an update is as reversible as a forget. revert puts a version back on the same memory - same id, so links to it still hold - and keeps what it replaced as a version in turn.

purge removes a memory's versions with it, and refuses to run without --yes. export writes current memories only; versions are not in an export, so a copy of the database file is the backup that carries them. All commands honor the same TETHER_DB/TETHER_SYNC_* env vars as the server.

Safety limits and privacy

Memory outlives the session that wrote it, so tether treats what it stores, and what it is asked, as untrusted:

  • Titles are one line. Line breaks and control characters collapse to a space, and a new memory's title is capped at 500 characters, because titles are printed in the index every session auto-loads; without this a saved title could forge further entries or section headers there. The index also shows at most 200 characters of any title.

  • Links take the ids of current memories only. Anything else (a word, a forgotten memory) is dropped and reported in dropped_links; a link from a memory to itself is ignored.

  • Requests are bounded: limit, budget, the number of tag filters and the number of distinct query words all have ceilings well above real use.

  • Errors meant for the agent are plain messages; anything internal is logged to the server's stderr and returned as a generic failure.

  • Files are private. A new database is created 0600, a directory tether creates 0700, and tether export -o writes a 0600 file. An existing database keeps its owner and group access and only loses access for everyone else; an existing directory is left as it is.

  • Sync refuses to send the token unencrypted across the internet: an http:// or ws:// sync URL is accepted for this machine, a private network address or a single-label host name. For anything else set TETHER_SYNC_ALLOW_INSECURE=1 (a VPN, say), or use libsql:///https://.

  • A bad request never turns sync off. Only a write that fails on the replica and then succeeds on a local connection does.

License

MIT

Available Tools

4 tools
forgetA

Soft-delete a memory by id: marks it no longer current (excluded from recall/the boot index) but keeps the row, reversibly, like consolidation and the forgetting sweep already do. Returns {"forgotten", "existed"}. (Permanent purge is an admin-only CLI operation, not available here.)

Inputs: id (required).

recallA

Search memories by keyword and semantic similarity, then follow the usage graph to related memories, most relevant first.

Ask in plain language, the way you would ask a colleague ("how do we run the integration tests?", "what did the user decide about the auth library?") - a memory matching some of the words is a hit, and the best match ranks first. Recall BEFORE starting a task, not only when stuck: the index you were given at session start is titles only. Memories from the current project rank slightly ahead of equally-good ones from elsewhere.

Each hit carries {id, type, title, body, tags, updated_at} and a via receipt explaining why it surfaced (a direct match, or the edge it was reached through). Use updated_at to judge staleness (an old fact may no longer hold; verify before relying on it) and id to cite what you update via remember/link.

body is an EXCERPT centered on your query, not the whole memory. When a memory was longer than the excerpt, the hit also carries truncated: true and body_chars (the full length). To read one in full, call recall(id=N) - that returns just that memory, whole.

Args: query: free text; punctuation is safe. May be omitted if tags is given. type: optional filter ("user"/"feedback"/"project"/"reference"). limit: max results (default 20). budget: how far to follow associations (0 = direct matches only). session: optional id grouping related recalls so they prime each other. tags: optional comma-separated tags; exact-match filter (a memory must carry every listed tag). Combine with query to filter its ranked hits, or use alone (query omitted) to list every current memory with those tags, newest first, deterministic rather than ranked - raise limit to fetch beyond the default page size. id: fetch this one memory in full instead of searching. Use it after a search returns a truncated hit you want to read completely. full: return complete bodies for every hit instead of excerpts. Costs the whole payload - prefer id= for the one memory you actually need.

Inputs: id, full, tags, type, +4 more.

rememberA

Save a durable memory. UPSERTS: a memory of the same type with the same (whitespace/case-normalized) title is updated in place instead of duplicated, so re-remembering a fact refines it rather than cluttering.

Worth remembering: decisions and their reasons, conventions, gotchas that cost time, the user's preferences and how they like to work, facts about the environment (paths, commands, accounts) you had to discover. Not worth it: anything derivable from the code, or a transcript of what you did. Remember it when you learn it, not at the end of the session.

Memories of type project/feedback/reference are tagged with the current project automatically (proj:<name>, from CLAUDE_PROJECT_DIR) unless you pass a proj: tag yourself; user memories are about the person and stay global.

Args: type: one of "user", "feedback", "project", "reference". title: a short label; also the dedup key within a type. body: the fact. For feedback/project, a "Why:" / "How to apply:" line helps. tags: optional comma-separated tags. links: optional list of related memory ids. Merged (union) into any links already on the memory, never replaces them - omitting this on a refine call preserves links set earlier. crystallizes: optional list of source memory ids this memory abstracts; links it over them as a crystallized principle (needs TETHER_CRYSTALLIZE).

Returns {"id", "action"} where action is "created", "updated", or (with TETHER_CONSOLIDATE on) "consolidated" - a near-duplicate was superseded.

Inputs: body (required), type (required), title (required), tags, +2 more.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.2/5.0

    Scored across 4 tools

    Disambiguation5/5

    Each tool has a clearly distinct role: remember writes/upserts, recall searches/reads, link creates relationships, and forget deletes. There is no overlap or ambiguity between the four operations.

    Naming Consistency5/5

    All tool names are single lowercase verbs that directly describe their action: remember, recall, link, forget. The naming pattern is uniform and predictable.

    Tool Count5/5

    Four tools cover the core memory operations without redundancy or bloat. This is a well-scoped set for a persistent memory server.

    Completeness5/5

    The set provides full lifecycle coverage: create/update via remember, read via recall, delete via forget, plus relation management via link. No essential memory operation is missing.

    Maintenance

    ActivityMaintained
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      F
      maintenance
      Local-first, auditable memory for AI agents. Provides durable context for MCP hosts with SQLite storage, CLI, and MCP tools for memory management.
      2
      Apache 2.0
    • A
      license
      C
      quality
      B
      maintenance
      Local-first, governable long-term memory for AI agents. Provides SQLite-backed storage, cross-session recall, and traceable memory corrections through a standard MCP interface.
      40
      1
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI agents and any MCP-compatible client to persist, search, update, and delete plain-text memories across sessions, using BM25 lexical ranking over SQLite FTS5 with namespace filtering, tags, free-form metadata, and JSONL export/import — entirely local, with zero external infrastructure.
      1
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables multiple AI agents and clients to share a self-hosted memory store over MCP or REST, supporting note creation, search, change tracking, and optional semantic retrieval from desktop, web, or mobile.
      MIT