plugmem-host

⚠️ Experimental. plugmem is mostly an AI-built experiment — written with the help of a small local model (Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf) and various Claude models, in roughly equal measure. Expect non-professional design choices, rough edges, broken behavior, or mistakes. Use it at your own risk.

plugmem-host is the std host layer for the plugmem temporal-memory engine. It supplies what the no_std engine does not own — files, locking, and network — so from this one crate a Rust program gets remember / recall / revise / forget plus graph link/unlink, backed by durable storage. It re-exports the engine, so this one crate is all a Rust program needs.

No embedding model is required. Of the four recall sources, only the vector one needs an embedder; text, graph and time work with nothing but the database. Unlike the CLI and the MCP server, a Rust caller has two ways to supply one — an [embedder] section in config.toml, or DatabaseBuilder::embedder — and a third option, passing a vector per call, which needs neither. See Embedding is optional below.

Which crate do I need?

Writing Rust and just want a working memory? This is the crate — it has everything. The others are for narrower needs.

Configuration

The shared config.toml loader and platform-aware database paths are documented in the full settings reference. plugmem-cli, plugmem-mcp, plugmem-napi and plugmem-py use the same settings catalogue and database-path precedence; only their explicit override syntax differs.

The config feature is enabled by default for this native host crate, so Settings::load and read_config are available without extra feature flags. Applications that construct Config programmatically and want the smallest host dependency surface can use default-features = false; this does not alter the plugmem-core or WASM build.

What recall does

Recall is not a vector lookup — it fuses four sources by reciprocal-rank fusion with a recency boost (tags filter; they are not a source):

Two clocks

Every fact carries two timestamps, and the distinction is the reason the temporal source exists at all:

• valid_from / valid_to — when the statement was true.
• recorded_at — when the memory learned it. Stamped by the engine.

One timestamp cannot hold both, and the host passes now explicitly, so the whole sequence is expressible:

# use plugmem_host::{Config, Database, RecallQuery, RememberInput};
# let (db, _) = Database::builder(Config::default()).open("agent.plugmem")?;
const JAN: u64 = 1_767_225_600_000;
const MAR_5: u64 = 1_772_668_800_000;
const MAR_10: u64 = 1_773_100_800_000;

// In January the memory learns where kim lives.
let first = db.remember(RememberInput {
    entity: Some("kim"),
    ..RememberInput::text(JAN, "lives in Moscow")
})?;

// On March 10th it learns the address changed. `revise` closes the old
// interval at this instant; it does not delete the record.
db.revise(
    first.id,
    RememberInput {
        entity: Some("kim"),
        ..RememberInput::text(MAR_10, "lives in Berlin")
    },
)?;

// Now: Berlin.
db.recall(RecallQuery {
    entities: &["kim"],
    ..RecallQuery::text(MAR_10, "kim")
})?;

// As of March 5th: Moscow. Still valid then, and already recorded.
db.recall(RecallQuery {
    entities: &["kim"],
    as_of: Some(MAR_5),
    ..RecallQuery::text(MAR_10, "kim")
})?;

// The graph depth belongs to this question: widen the walk for a neighbourhood
// query, or use `Some(0)` when only kim's own facts should answer.
let _around_kim = db.recall(RecallQuery {
    entities: &["kim"],
    graph_depth: Some(3),
    ..RecallQuery::text(MAR_10, "kim")
})?;
# Ok::<(), plugmem_host::HostError>(())

as_of moves both clocks: a fact answers only if it was valid at that instant and had already been recorded by then. The second half is the one people trip over. An as_of earlier than a fact's recorded_at does not see it, because the memory genuinely knew nothing then — answering with today's knowledge would be the wrong answer to "what did I hold".

valid_from is the other half of the model: a statement that became true before you heard of it. Recording on March 10th that the move happened on March 1st — valid_from: Some(MAR_1) on the revision — closes the old interval at March 1st instead of at March 10th. A query as of March 5th then finds neither: Moscow had stopped being true, and Berlin was not yet known. That is not a hole in the model, it is the honest answer for that instant, and it is exactly what one timestamp cannot express.

The old record stays on disk either way. forget is the destructive verb: use it when a fact was wrong, not when it changed.

Edges are temporal the same way, so an as_of traversal walks the graph as it stood then — through relationships that have since been unlinked.

What plugmem-host adds

The retrieval above lives in the engine; this crate adds the OS side:

• File-backed storage — atomic snapshots (tmp + fsync + rename), an append-only journal with a configurable fsync policy, crash recovery (a torn journal tail is detected and dropped on open);
• OS locking, one writer — an advisory lock per database file. A second read-write open is refused with a typed HostError::Locked rather than corrupting the file silently. Readers do not take that lock;
• A read-only mmap open, concurrent with the writer — Database::open_readonly pins the current published snapshot generation and maps it, borrowing the mapped pages so a large read-mostly database resides only in the pages a query touches. Generations are immutable — a checkpoint writes a new one and never rewrites an old one — so a writer and any number of readers run at the same time, across threads and across processes, sharing the OS page cache. This is the WAL/MVCC arrangement, not "N readers or one writer": a reader sees state as of the checkpoint it pinned (snapshot isolation) and calls refresh() to adopt a newer one;
• A write path that does not clone the file — Database::open memory-maps the snapshot and the engine borrows it as an overlay: mutations land in a small owned overlay (an appended tail plus per-page copy-on-write), so opening a multi-gigabyte database to append one fact no longer copies the whole image into RAM. A snapshot materializes the base + overlay into a fresh file and re-maps it. Validation is lazy (the SQLite model): an open range-checks the record metadata and nothing else, so the large text, vector and per-fact-metadata pools stay non-resident until a query touches them — a measured open residents well under half of a text-heavy image. What that guarantees is precise and worth stating: no stored id can make a read unsafe. It does not mean the data agrees with itself. The default open never checksums the image either, so it stays sparse; corruption is caught when the bad record is read (never a panic), or on demand with verify() (content and graph consistency) and scrub() (byte-level container integrity) — see Integrity & recovery below;
• Maintenance policy — auto-snapshot and optional auto-maintain, run inline (no background threads);
• Embedding providers — one HTTP client for the /v1/embeddings shape shared by OpenAI, Ollama, LM Studio, vLLM and llama.cpp-server.

Without an embedder it is fully functional — lexical, tags, graph and time still answer; vectors are an addition, not a requirement.

Example

use plugmem_host::{Config, Database, OpenAiCompatEmbedder, RecallQuery, RememberInput};

let mut cfg = Config::default();
cfg.dim = 768;

let (db, _report) = Database::builder(cfg)
    // one client covers OpenAI, Ollama, LM Studio, vLLM, llama.cpp-server:
    .embedder(Box::new(OpenAiCompatEmbedder::new(
        "http://localhost:11434/v1", // Ollama's OpenAI-compatible endpoint
        "nomic-embed-text",
        768,
    )))
    .open("agent.plugmem")?;

// The text is embedded automatically (outside the database lock),
// quantized and indexed:
db.remember(RememberInput::text(1_784_000_000_000, "prefers tokio"))?;

// Bulk load: the whole batch is embedded in one round-trip and the
// journal is fsynced once (not per fact) — the write path for `import`:
db.remember_many(vec![
    RememberInput::text(1_784_000_050_000, "uses pinned versions"),
    RememberInput::text(1_784_000_060_000, "lives in Berlin"),
])?;

// The query text is embedded too; recall fuses lexical, vector, graph
// and temporal evidence into one ranked, token-budgeted block:
let out = db.recall(RecallQuery::text(1_784_000_100_000, "which runtime?"))?;
println!("{}", out.rendered);
# Ok::<(), plugmem_host::HostError>(())

Benchmarks

cargo run -p plugmem-host --example edge_lifecycle
cargo run -p plugmem-host --example maintain_modes
cargo run --release -p plugmem-host --example bench_database -- 100000 --diagnose-recall | tee database-benchmark-100k.tsv
cargo run --release -p plugmem-host --example bench_database -- 1000000 --diagnose-recall | tee database-benchmark-1m.tsv
cat database-benchmark-100k.tsv database-benchmark-1m.tsv > database-benchmark-scale.tsv
cargo run -p plugmem-bench-charts -- database-benchmark-scale.tsv --force
cargo run --release -p plugmem-host --example bench_edges -- 100000 | tee edge-benchmark-100k.tsv
cargo run --release -p plugmem-host --example bench_edges -- 1000000 | tee edge-benchmark-1m.tsv
cat edge-benchmark-100k.tsv edge-benchmark-1m.tsv > edge-benchmark-scale.tsv
cargo run -p plugmem-bench-charts -- edge-benchmark-scale.tsv --force
cargo bench -p plugmem-host

The committed assets/database-*.svg charts are generated from the database runs; assets/edge-lifecycle-*.svg charts are generated from bench_edges.

For the same-workload comparison between 100k and 1M, see the measured scale table on GitHub. For the public Rust API, see the plugmem-host documentation on docs.rs. The database example uses deterministic synthetic facts and dim=0; it does not call an embedding service or use a network connection.

Native builds are 64-bit, so a host process reads every capacity class of the shared file format: databases sized for the 32-bit wasm budget (≤ 2 GiB, the default) and databases with larger limits alike. Opening, reading, scanning and checkpointing go through the mmap overlay, whose clean pages the OS can reclaim — so those work on a database larger than RAM. A rebuild (maintain and recover) is disk-first too: it streams the two big pools (vectors, text) through temp files and keeps only the metadata and the HNSW graph resident, so peak RAM tracks the record count, not the image size. The residual limit is the graph itself — a database whose graph exceeds RAM is a further tier. The per-structure byte costs and pool limits (what a fact, an edge or a vector weighs, and where each tops out) are tabulated in plugmem-core's Capacity — what weighs what. The snapshot format is pointer-width independent — a file written here opens unchanged in a wasm32 or wasm64 build of the core, as long as its configured limits fit that host.

Files

One local database rooted at agent.plugmem consists of:

Snapshot writes are atomic (tmp + fsync + rename + directory fsync): a reader observes the old image or the new one, never a torn file. Journal appends are fsynced per operation by default (FsyncPolicy::EachOp); OnSnapshot trades the crash-window of the journal tail for speed. On open, the journal is replayed over the snapshot deterministically; a torn tail from a crash mid-append is detected, dropped and reported.

Workspaces (optional)

Default: one Database, one local database layout. Workspace is for a process serving many independent memories — a directory of named databases, opened on demand and pooled, plus an optional registry of what each is for.

use plugmem_host::{DbName, IfMissing, Settings};

let ws = Settings::load(None)?.open_workspace("/srv/bot".as_ref())?;
let db = ws.get(&DbName::parse("chat-42")?, now_ms(), IfMissing::Create)?;
# fn now_ms() -> u64 { 0 }
# Ok::<(), Box<dyn std::error::Error>>(())

A name is [a-z0-9][a-z0-9_-]* and cannot represent a path, so it resolves to exactly one named database inside the directory. The pool bounds how many stay open; close_idle releases the rest, which matters because an open writer holds the file's exclusive lock — the timeout is a liveness setting, not a memory one.

The core is untouched by any of this: one Memory is still one database. See specs/10-workspace.md.

Concurrency model

The engine is single-writer by design. The host runs a WAL/MVCC-style versioned layout, so one writer and any number of readers run at the same time — across threads and processes — without a reader ever blocking the writer or seeing a torn image.

• One writer. Database::open takes an exclusive writer lock; a second writer gets HostError::Locked immediately (a typed refusal, not silent corruption). The lock dies with the process, even on a crash. Readers do not take this lock, so they never contend with the writer.
• Many readers, concurrent with the writer. A checkpoint never overwrites a live file: it writes a new immutable snapshot generation and repoints a tiny manifest. Database::open_readonly pins the current generation with a shared lock and maps it — so it coexists with a live writer and reads a consistent snapshot "as of the last checkpoint" (it does not see writes made after it opened; reopen to advance). Readers across threads or processes share one copy of a generation in the OS page cache. The writer reclaims a superseded generation only once no reader still pins it, so disk stays bounded by the longest-lived reader.
• One process, many threads or agents. A Database is a Clone + Send + Sync handle; clone it freely. The read verbs (recall/get/stats/export/verify) take a shared guard and run concurrently; the write verbs take an exclusive guard and serialize — against each other and against readers (an RwLock over the engine, the same reader/writer discipline as the file lock, one level down). At microsecond engine calls neither is a bottleneck. ReadOnlyDatabase is Send + Sync too, and its reads are lock-free — a fan-out of reader threads over one mapped snapshot. (Each reader thread keeps its own recall scratch, so concurrent recalls never contend.)
• Many files. Fully independent databases: separate locks, separate mutexes, natural parallelism. Two models each with their own memory file never contend; two models sharing one memory clone one handle.
• Network stays outside the lock. Embedding calls (the slow, external part) run before the mutex is taken, so an agent waiting on its embedding provider does not stall the others.

The typical shape is build occasionally, read a lot: one writer snapshots the memory (on a schedule or in a maintenance window), then many read-only consumers query it in parallel.

use plugmem_host::{Config, Database, RecallQuery};

// Many readers over one checkpointed file — zero-copy, shared page cache.
// (A read-only open needs a published generation: checkpoint once, first. It
// then reads that generation and ignores any later journal, so a reader sees
// the memory as of the last checkpoint.)
let ro = Database::open_readonly("agent.plugmem", Config::default())?;
let out = ro.recall(RecallQuery::text(1_784_000_100_000, "which runtime?"))?;
println!("{}", out.rendered);
# Ok::<(), plugmem_host::HostError>(())

Memory-mapped opens and disk-first rebuilds

Both open and open_readonly memory-map the snapshot; the engine borrows an overlay over the mapping rather than reading it into a heap copy. Opening is an mmap plus a bounds-check regardless of file size; only dereferenced pages fault in, and the OS may evict clean ones. Resident memory tracks the working set, not the file size.

A read-write handle appends new records to a small owned tail; the mapped base is never rewritten in place (the append-only structures avoid copy-on-write). A checkpoint streams the fresh image to a temp file and renames it atomically, unmapping first (Windows will not rename a mapped file).

Read-only handles take a shared lock and map the same snapshot, so several share one copy in the OS page cache. Open read-write only to mutate.

maintain and recover run disk-first: the two large pools (vectors, text) stream through a temp Scratch file (a sibling of the database, mapped on freeze, deleted on drop), keeping only metadata and the HNSW graph resident. Peak RAM is proportional to record count, not content size. On no_std (no files) the engine uses the in-RAM rebuild instead.

Integrity & recovery

The default open trusts the file (like SQLite): it does not checksum the whole image, so a large database opens sparse. Integrity is on demand, in three layers of increasing cost, and corruption is never a panic — the accessors tolerate bad bytes, these turn latent damage into an explicit error or a repaired file.

scrub() — the bitrot detector. A resumable iterator over the mapped snapshot: each next() hashes up to a slice budget, so you pace it yourself (run to completion, or a slice at a time on a background thread, pausing/cancelling between slices). It holds a shared lock for its whole life, reads the map linearly (pages fault in, get hashed, stay reclaimable — it never residents the whole file), and reports the first mismatch, naming the damaged section.

A scrub reads the file, not a handle's view of it, so it is on both handles. Use the writer's when you already hold one: a second read-only open would map the whole image again and take a second lock to hash bytes whose path you already know.

use plugmem_host::{Config, Database};

let (db, _) = Database::open("agent.plugmem", Config::default())?;
// Verify every container byte, a slice at a time.
for step in db.scrub()? {
    let progress = step?; // Err(Corrupt) names the first damaged section
    // e.g. report progress.done_bytes / progress.total_bytes to a UI
    let _ = progress;
}
# Ok::<(), plugmem_host::HostError>(())

It needs a published generation — checkpoint once — but not a clean journal: it checks the container as it stands, and the journal describes the generation that has not been written yet.

recover() — Tier 2 salvage. For content corruption (bad text bytes, a broken vector bijection): it opens the source, drops the facts that fail verify()'s per-fact checks, compacts the survivors and their indexes, and writes a fresh image to a new file — leaving the source untouched as evidence. It returns a RecoverReport { kept, dropped_text, dropped_vector }. It is disk-first (the source opens as an mmap overlay and the two big pools stream through temp files), so peak RAM tracks the record count — a database larger than RAM can be recovered, as long as its graph fits.

use plugmem_host::{Config, Database};

// now = a millisecond timestamp; dst must differ from src.
let report = Database::recover("agent.plugmem", "agent.recovered.plugmem",
                               Config::default(), 1_784_000_000_000)?;
println!("kept {}, dropped {} text + {} vector",
         report.kept, report.dropped_text, report.dropped_vector);
# Ok::<(), plugmem_host::HostError>(())

What recover does not do. Structural damage — a snapshot that will not even parse — is not salvageable here: the source fails to open and recover returns the typed error; restore from a backup instead (Tier 0).

Recovery layers (first release). Most recovery is not salvage at all:

• Tier 0 — restore. A snapshot is one atomic file (tmp + fsync + rename); back it up and copy it back. scrub() tells you when to. This covers the overwhelming majority of cases.
• Tier 1 — regenerate. Re-ingest from your upstream source (logs, documents) into a fresh database.
• Tier 2 — recover(). Content-corruption salvage, as above.

Maintenance policy

Configured through the builder, executed inside the same critical section as the operation that triggered it — there are no background threads, matching the engine's own philosophy:

maintain is policy-driven. The default Auto path first checks whether anything is pending; with no tombstones, stale text index or vector tail to optimize, it returns a no-op report without rewriting the snapshot. When work is pending, host maintenance stays disk-first: text and vector pools stream through scratch files, ordinary BM25 compaction filters existing postings, and HNSW work is bounded unless a full rebuild is explicitly requested.

maintain_with_options selects the policy explicitly. No mode ever drops a fact revision or an edge version — what the heavier modes buy is bytes and index freshness, never less history.

Full is the only mode that repacks the edge arenas. Relinking many relations fragments them — the incoming mirror is keyed by the far endpoint, so interleaved runs keep splitting pages in half — and rewriting them in key order packs the pages again. Measured over 200 relations relinked 1000 times (200k retained versions): 31.9 MB → 23.4 MB, 59 ms, every version kept.

The same selection is exposed by plugmem maintain --mode <mode>, the mode argument of the MCP plugmem_maintain tool, and maintain(mode?) in the Node bindings.

Embedders

The Embedder trait is two methods (dim, batched embed). OpenAiCompatEmbedder speaks the /v1/embeddings shape that OpenAI, Ollama, LM Studio, vLLM and llama.cpp-server share — there is no provider-specific client because there is no provider-specific protocol. The dimension is configured explicitly (no startup probe); a server answering with a different one is a typed error. Tests run against a local mock — no network in CI.

embed takes &self, and the trait requires Send + Sync, because an embedder is a client of a remote service rather than a piece of mutable state. That is what lets every caller share one instance with no lock in front of it: SharedEmbedder is a plain refcount, a Database holds its embedder unlocked, and concurrent verbs sit in the provider at the same time instead of queueing behind one HTTP request. An implementation that does need mutable state (a cache, a rate-limit budget) brings its own interior mutability, which is the only place that knows what may overlap.

Feature flags

• serde — Serialize/Deserialize on the public data types (FactSnapshot, ExportedFact, RecoverReport, FsyncPolicy), forwarding to plugmem-core/serde. Off by default. HostError is deliberately not covered — it wraps std::io::Error, which is not serializable.
• counters — deterministic work counters for the perf gates, forwarded to plugmem-core/counters. Do not enable it in normal use — it is a single-threaded measurement build only. The arena's counter Cells are not Sync, so with counters on the engine lock falls back from an RwLock to a Mutex: the read verbs then serialize instead of running concurrently. The public API is identical either way — only the internal lock (and thus read concurrency) changes. Leave it off to keep concurrent readers; reach for it only when measuring operation counts in a specific scenario.

License

MIT.