mimir-core 0.1.0

Mimir — agent-first memory core: foundational types and invariants.
Documentation

mimir_core

Foundational crate of Mimir, an experimental pre-1.0 memory governance system for AI agents. mimir_core provides the librarian's deterministic core: a compiler-pipeline (lex → parse → bind → semantic → emit) over a Lisp-shaped agent-native IR, plus the bi-temporal append-only canonical store, plus the read-side query engine.

Pre-1.0 status. This crate is part of Mimir's active-development tree. APIs, storage details, and wire-format compatibility may change before v1. Public crates.io releases wait for the first alpha.

Install

Until the first alpha release, use a workspace or path dependency:

[dependencies]
mimir-core = { path = "/path/to/Mimir/crates/mimir-core" }

Quickstart

use mimir_core::{Store, Pipeline};

# fn main() -> anyhow::Result<()> {
let path = tempfile::tempdir()?.path().join("workspace.log");
let mut store = Store::open(&path)?;

// Commit a batch of agent-native memory forms.
let now = mimir_core::ClockTime::from_millis(1_700_000_000_000);
store.commit_batch("(sem @alice :p @likes :o @rust :c 0.95)", now)?;

// Query it back.
let result = store.pipeline_mut().execute_query(
    "(query :kind sem :s @alice :p @likes)",
    now,
)?;
assert_eq!(result.records.len(), 1);
# Ok(())
# }

What's in here

  • Compiler pipeline: lex, parse, bind, semantic, pipeline::Pipeline::compile_batch — agent input → typed canonical records.
  • Canonical wire format: canonical — [opcode][varint length][body] framing, 18 opcodes, fully self-describing, fuzz-target-covered.
  • Append-only store: Store over a LogBackend-abstracted CanonicalLog. Two-phase commit with crash-injection-tested rollback.
  • Read-side: Pipeline::execute_query over the in-memory current-state index; p50 ≈ 0.57 µs on a 1 M-memory warm index.
  • Decay: integer-fixed-point exponential decay via a hand-baked 256-entry lookup table — bit-identical across architectures.
  • Workspace partitioning: WorkspaceId = hash(git_remote_url) with an in-process .git/config parser. The current implementation keeps raw workspace memories isolated; the draft scope model adds governed promotion above this layer.

Engineering posture

  • #![forbid(unsafe_code)] workspace-wide.
  • Full test suite: unit + property + doctest + integration + crash-injection + 3 fuzz targets (live counts in STATUS.md frontmatter).
  • 19 thiserror-derived error enums — one per subsystem.
  • cargo deny check gates licenses + advisories + sources on every PR.
  • unwrap_used = "deny", expect_used = "deny", dbg_macro = "deny" workspace-wide; relaxed only inside #[cfg(test)].

Specs

The original 14 architecture specs in docs/concepts/ are all authoritative; scope-model.md is a draft mandate-expansion spec. Every load-bearing claim in this crate cites the spec section that backs it via module-level //! docs.

License

Apache-2.0.