v0.1.0 — initial public release

Changelog

All notable changes to this project are documented here.

The format follows Keep a Changelog,
and Flashback adheres to Semantic Versioning.

0.1.0 — 2026-05-23

Initial public release. Flashback is a self-contained Rust microservice that
gives any LLM dynamic, episodic memory: a four-tier hierarchy with append-with-
supersede history, real-time within-conversation ingest, and hybrid retrieval
over a temporal graph backed by pgvector.

What works in 0.1.0

Memory model

• Four-tier hierarchy: core (always-injected) / working (TTL'd) / episodic /
  semantic.
• Append-with-supersede — old memories never deleted; superseded rows stay in
  the lineage chain for /lineage queries.
• Default retrieval returns the terminal node; lineage traversal exposes the
  full evolution.

Ingest

• POST /memory/ingest accepts raw text or structured user/assistant turn
  pairs.
• Pluggable AiProvider for extraction:
  • heuristic — rule-based, in-process, zero network. Default.
  • remote — any OpenAI-compatible HTTP endpoint (OpenRouter, Anthropic,
    OpenAI, local Ollama, etc.).
  • embedded — LLM running in-process via mistralrs (air-gapped / single-
    box deploys).
• Per-role model tiering: separate extract (fast, ~2s budget) and distill
  (background, minutes-OK) models — see docs/MODEL-TIERING.md.

Retrieval

• POST /memory/search — hybrid: vector cosine + BM25 keyword + recency +
  project-match + entity overlap.
• answer mode (relevance-weighted) and manager mode (situational-
  awareness-weighted).
• POST /context/assemble — structured 5-layer prompt: procedural / active-
  project / retrieved-memories / document-chunks / recent-conversation.

State objects

• Typed mutable state (/state/{kind}/{key}) with op-based patches and
  full supersede history. todo_list is the first shipped kind.

Auth

• Bearer-token, sha256-hashed at rest, scoped per user.
• Plaintext shown once at mint; flashback token mint --user=<user> --name=<label>.
• --dev / FLASHBACK_DEV_MODE=1 bypasses auth for local development;
  banner-warned on every startup.

MCP transport

• Streamable-HTTP MCP server on :8082/mcp, wraps the REST API as typed
  tools.
• Wire into Claude Desktop / Cursor / Claude Code by pasting the URL + bearer
  into the client config — see README.

Storage

• Postgres + pgvector 0.4.2 (with SQLx 0.9 support).
• sqlx::migrate! baked-in migrations; AUTO_MIGRATE=1 runs them on first
  boot.
• fastembed-rs for embeddings; ONNX model cached at
  /opt/flashback/fastembed-cache (pre-fetched at Docker build time).

Deploy

• docker compose up — Postgres + sidecar + REST + MCP, all wired.
• DigitalOcean App Platform spec at .do/app.yaml — one-click deploy.
• deploy/install.sh for fresh-VPS bootstrap.

Consolidation

• Background worker promotes working → episodic and distills episodic →
  semantic on configurable intervals (daily + weekly defaults).
• Per-user scoping; results logged in consolidation_runs.

Security posture (shipped with 0.1.0)

CI / repo hygiene running on every push + PR:

• cargo fmt --check, cargo clippy, cargo test, release build (ci.yml)
• SonarCloud scan with cargo llvm-cov coverage (build.yml)
• actionlint, trufflehog --only-verified, gitleaks detect,
  cargo deny check (advisories + bans + licenses + sources),
  semgrep --config auto, actions/dependency-review-action on PRs
  (security.yml)
• GitHub CodeQL with the security-and-quality query suite for Rust
  (codeql.yml)
• OpenSSF Scorecard, weekly + on push, publishes public score
  (scorecard.yml)
• CycloneDX + SPDX SBOMs generated and attached on every release
  (sbom.yml)
• Dependabot with grouped major + minor-and-patch, capped at 5 PRs per
  ecosystem per week.

Repo-level:

• Secret scanning + push protection on (GitHub-native).
• Dependabot vulnerability alerts + automated security updates on.
• All GitHub Actions SHA-pinned (except ossf/scorecard-action, which the
  Scorecard webapp requires as a tag pin — documented inline).
• CODEOWNERS, PR template with security checklist, bug + security-contact
  issue templates.

Not in 0.1.0 (designs documented; implementation deferred)

• Multi-tenant isolation. docs/TENANCY.md is exploratory — visibility
  scoping, group memberships, per-tenant consolidation are designed but not
  shipped. Today every memory belongs to one user_id and admin endpoints
  see all of a user's memories.
• Cognitive modes. docs/MODES.md is exploratory — per-project default
  modes, caller overrides, LLM auto-classification all designed but not
  shipped.
• Branch protection on main. Solo-dev project today; deliberate
  trade-off until contributors join.
• Custom secret-scanning patterns (non-provider + validity-checks).
  Require an org-level toggle on Horizon-Digital-Engineering; basic
  GitHub-provider scanning is on.
• Private vulnerability reporting. Org-level toggle; not configured.

License

Business Source License 1.1. The Licensed Work is © 2026 Horizon Digital
Engineering LLC. Non-production use is freely permitted. Production use is
permitted except for offering Flashback as a hosted or managed service that
competes with one offered by the Licensor. License auto-converts to Apache
License 2.0 on 2030-05-23 (four years from this release).

See LICENSE for the full text.