MemophantMemophant
Product

Every release, in plain language.

What's new, improved, and fixed in each version of Memophant. Updates arrive automatically in the app — this is the human-readable story of what changed.

  1. 2.1.1

    A stability release — a smoother, more responsive app and a memory engine that holds up under heavy use.

    Improved

    • Fewer hitches. The Sessions tier, the Code Browser, drift checks, tier lists, repository scans, and previews now do their disk and database work off the main thread, and a Code Browser search you've typed past is cancelled instead of finishing in the background.
    • Steadier under bursts of writes. Background follow-up work after each memory write is now capped, so a large batch of changes commits several times faster without piling up database connections.

    Fixed

    • Switching notes or projects quickly could let a slow result land in the wrong place — views and editors are now tied to the item they were opened for.
    • The MCP tool-call timeout now actually fires, and a tool call that arrives during a license unlock is completed instead of silently dropped.
    • Moving a memory while it was being edited could lose the edit. The move now keeps it, or fails cleanly with nothing lost.
    • Several rare search-index races that could briefly show stale results or attach the wrong embedding to a note.
    • The code auto-indexer starts and stops cleanly, and overlapping project refreshes are merged rather than dropped.
    • A failed model load could trigger a duplicate ~80 MB download of the embedding model.

  2. 2.1

    Your project's memory now largely maintains itself — and every number, count, and progress message the app shows you about that work is honest.

    New

    • The Memory Review window. Reviewing flagged memories is no longer a cramped sheet — it's a dedicated, resizable window per project. Every item carries the evidence you need to make the call: real before/after diffs for proposed changes, both notes side by side for conflicts, masked context when a secret was blocked, and a banner when a note changed since it was flagged. The action buttons stay pinned in view no matter how long the note is.
    • Charter drafting with Claude. A project without a charter can have Claude draft one — from the editor, or automatically in the background at a pace you control. Drafts land in a review pen where you accept, edit, or discard; you remain the only one who can save the actual charter, and every write is secret-scanned first.
    • The autonomous Steward, on for everyone. Background memory curation and trusted auto-apply are now enabled by default — including for existing installs (Settings turns it off). The Steward fixes what it can, escalates what genuinely needs you, and every change it applies is one click to revert.

    Improved

    • Counts you can trust. The "Review N" button, the health card, and the review window now count exactly the same things — the number on a button is the rows you'll see when you click it. Items the Steward gave up on are finally reviewable, and clearable.
    • Honest progress. A running drain shows what this run will actually do rather than the whole queue; spot-checks and section consolidation report their own phases instead of freezing on stale numbers; and the history tab's quality signal now measures challenges against the fixes they challenge — honestly re-verifying a correct note no longer reads as lost trust.
    • Faster curation. One corpus scan per drain instead of one per note, related notes batched into clusters with a single verification pass, and drift evidence resolved up front so investigations start warm.
    • An accurate activity feed. Drift rows stop re-announcing themselves on every launch and no longer count the same commits once per tier; the Steward's edits are labeled as its own instead of blending in with yours.

    Fixed

    • Reverting an auto-applied fix now correctly reduces the Steward's trust in that kind of change — trust can go down, not just up.
    • The review button could read "Review 39" over an empty list; the invisible backlog behind it is now garbage-collected and everything that remains has a surface.
    • Detached (browse-only) projects can no longer be written to by any drafting or curation path.
    • Hand-edited agent instruction files with ambiguous markers are reported plainly instead of being silently skipped forever.
    • A pass of copy fixes across the dashboard so no surface claims more than the app actually did.

  3. 1.5

    Your project's memory now follows your project — and everywhere it can't follow, it says so out loud instead of quietly doing the wrong thing.

    New

    • Folder renames now follow. Rename or move a project folder on disk and Memophant notices and offers to migrate it — carrying the search index, review stamps and access history, note anchors, permalinks, and the project's stored credentials across. The migration is resumable and safe to interrupt; if it can't finish, it rolls back and tells you rather than leaving you half-moved.
    • Anonymous usage statistics, with a real off switch. Memophant now reports anonymous, aggregate usage so development can follow what people actually use. It's privacy-by-construction: every event is a fixed list of names and counts, so there is no code path that could send a file path, a note's contents, a project name, or anything you wrote — that isn't a policy, it's a compile error. It's on by default, and Settings → Privacy turns it off. We verified against a live network listener that off means zero bytes on the wire — including the analytics SDK's own automatic events — that it stays off across restarts, and that re-enabling never replays what was suppressed.
    • If you're updating from 1.4, this release also brings you the Project Charter, per-hit secret-scan acknowledgments, and the sealed commit door — which shipped in between, in releases aimed at applications that embed the Memophant engine.

    Improved

    • A moved folder never fails silently. If a project's folder had been moved or deleted while Memophant still held its old path, writes and searches used to quietly recreate that dead folder and carry on — notes landing at a path you'd abandoned, searches answering "no results" out of an empty index they had just created. Now every door checks first: reads report that the folder is gone, writes refuse and name the problem, document import tells you instead of discarding the file, and the Code Browser shows the moved-folder message instead of "no symbols." A healthy project that has simply never been indexed still sets itself up on first use, exactly as before.
    • The MCP health check can no longer hang. On first launch, macOS Gatekeeper can hold the server just long enough that the old check waited forever — leaving the status pill stuck on "Checking…" and Reinstall permanently greyed out. The check now honors its own timeout.

    Fixed

    • Renames can't lose history. A rename interrupted and retried could wipe the review stamps, anchors, and access counts a previous pass had just moved — the one thing in a rename that git can't restore. Fixed, along with a guard so two overlapping renames can't strand a project between names.
    • Adopting an existing project keeps its detected name — a project already named in your MCP config could block the review button on its own name.
    • Usage-statistics accuracy: no more double-counted error events, bulk prune reports what it actually applied, and events can't reorder or drop when you quit.
    • A full pre-release audit of this span produced roughly seventeen hardening commits, including test repairs that had been silently passing since the last toolchain update. The engine test suite is fully green for the first time.

  4. 1.4

    Every path into your memory now runs through one gated engine — and that engine is now a product other apps can embed.

    New

    • The memory engine is now a real Swift package with a versioned public contract, so another application can link it in-process and get the same memory — same guards, same index, same secret scan — without running Memophant. The first external embed is already live and verified end-to-end.
    • Every route to durable state — memory, wiki, design, code, tasks, sessions, templates, vendors, and documents — now passes through the engine's gate: validation, the write-time secret scan, provenance, and indexing apply uniformly whether the write comes from the app, an agent, or the Curator. An enforcement instrument holds the number of bypass paths at zero so one can't quietly reappear.
    • One-shot search from the command line: memophant-mcp search answers a lexical query straight from the index with no server running, opening the index read-only so a query can never mutate your memory.
    • Claude hook wiring moved to user scope — registered once, machine-wide, instead of per-project. Existing projects migrate automatically, and a hook file that can't be written fails loudly instead of silently.

    Improved

    • Your writes can't be silently lost: every write funnel now rides a real compare-and-swap with recompose-on-conflict, so two writers landing on the same note merge instead of one clobbering the other — and editing a note refuses an ambiguous replace instead of guessing.
    • The secret scan got sharper with its wider reach: agent-written documents are scanned both when written and when imported, prose that merely talks about key formats no longer blocks a save, and no surface in the app will render a blocklisted secret.
    • Tasks joined the same argument contract as memory: creating a task refuses an invalid status or priority instead of silently swallowing it, and mistyped arguments are refused with the valid ones named.
    • A locked license session now recovers on the next call instead of staying dead until restart, and a locked gate reports the real reason it's locked.
    • The app stops guessing: a failed credential read shows as a failure rather than an empty state, and document import failures land in a report channel a reload can't clear.

    Fixed

    • Long subprocess output is no longer silently truncated, which could drop the tail of a generation or audit.
    • Spurious "database is locked" failures under concurrent access are fixed — the busy timeout now applies to every connection.
    • Renaming a note takes its permalink with it, fixing the ghost that broke editing a wiki page after a rename.
    • Custom memory folder locations are retired: memory is pinned to .memory/, and legacy configurations migrate automatically.
    • A project whose index has no rows now reports that condition instead of answering every search with an empty success.

  5. 1.3

    The biggest release since launch: a rebuilt engine underneath your memory, the same guards and drift-checking across every tier, and notes that start linking themselves as they're written.

    New

    • Memory runs on a rebuilt, database-authoritative engine — with your markdown still canonical on disk. Saving a note is now a single transaction with the embedding work moved off the critical path, so writing memory no longer stalls the agent doing it.
    • Search reaches one hop out through your notes' relations, so asking about a decision also surfaces what it builds on — and building context for a topic now pulls from your wiki, design, and code notes, not just memory.
    • Notes can anchor to a specific function or type instead of a whole file, so drift-checking gets precise — and any note documenting a symbol is reachable from the code side.
    • Your wiki, design, and code tiers now write through the same funnel memory always used: validation, the secret scan, provenance stamping, and immediate indexing. Wiki pages carry provenance and get surgically drift-checked, and curated pages can hold generated sections that refresh without touching your prose around them.
    • Writing a note tells you what it should connect to. Save something close to an existing note and Memophant surfaces the neighbors it found, with a single call that links them — validating every target so a link can never point at a note that doesn't exist. Notes that never got linked become work the Curator picks up on its own.
    • When an agent reads a note whose anchored code has drifted, the note now arrives carrying that warning and the one call that clears it — so a stale note gets verified instead of quietly trusted.

    Improved

    • The Curator can now graduate to applying a kind of change on its own — but only per repository, and only after it has been right ten times in a row with no reverts. A single revert demotes it back to proposing.
    • It converges faster, too: a backfill lane for missing anchors, tier canonicalization, and successor rebinding clear the busywork that used to accumulate between passes.
    • Corpus health answers in a readable summary by default — counts by cause, plus the worst offenders — and can be scoped to just the notes a session touched.
    • Code search is cleaner: generated and minified files no longer crowd out your own code, and an exact symbol-name match now ranks first.
    • A leaner agent surface — two tools nothing actually used were retired, leaving 25 — and every list-style tool now rejects a mistyped argument with a message naming the valid ones instead of silently ignoring it.
    • Multi-agent crews get scoped tool sets and crew-aware instructions, so a coordinating agent and a worker agent can see different surfaces of the same memory.
    • Retiring a note hides it from your live corpus without erasing it from history — retired notes stay queryable.

    Fixed

    • Prose that merely mentions link syntax is no longer mistaken for a real link — in the write-time nudges or in the Curator's own detection.
    • A note anchored to code whose symbol index failed to build no longer reports as having no symbols; a lookup failure and a genuinely empty result are now told apart.
    • Renaming a tier folder to fix its casing is no longer blocked by prose mentions of the old name in your notes, wiki pages, or task board.
    • Notes promoted into history are filtered correctly by validity, and page anchors backfill for pages that predated them.

  6. 1.2.5

    Your memory maintains itself in the background with the new Curator — and every Claude flow now runs on your Claude subscription, no API key required.

    New

    • The Curator: a background lane that continuously re-checks your notes against the code they describe, investigates real drift, and proposes a fix, re-ground, or retirement — applying the confident, low-risk ones itself and sending the rest to your review inbox. Every automatic change is journaled and revertible.
    • Every AI flow — wiki, code, and design generation, Migration, task Plan and Run, Consolidate, and Extract Templates — now runs on your Claude subscription through the Claude CLI. There's no Anthropic API key to paste, store, or meter.
    • One-click reconnect, with your Claude sign-in state visible across the dashboard, Settings, and every wizard; a failed run offers Try Again and a Copy Diagnostics report.
    • Agent-generated plans, reports, and briefs get a dedicated documents tier, fenced off from your project's own docs/, with a drift audit that offers to move anything misfiled.

    Improved

    • Grounding anchors more precisely — it resolves partial paths and follows a symbol to its definition site, so a note about a function anchors where that function actually lives.
    • The review inbox is sharper: proposals are a first-class filter, an emptied filter never strands you, and acting on an item no longer swaps the list for a spinner.
    • Tier folders that drifted in case (Documents/ vs documents/) auto-repair, and notes filed outside the six canonical folders can be refiled with a single confirm.

    Fixed

    • A secret — an API token or key — in a note is now blocked at write time, not just flagged afterward.
    • Licensed installs re-authenticate far less often: MCP access tokens now last 14 days for licensed users (24 hours on trial).
    • Notes created during Consolidation now pass through the same write guards as every other note.

  7. 1.2.1

    A point release that hardens Claude-powered generation and audit, and fixes a few sharp edges from 1.2.

    Fixed

    • Generating a wiki, code, or design page — or running a Migration — on a large repo no longer times out; the time budget now scales with how much the run has to work through.
    • Fixed the “model's response was not valid JSON” error that could end a generation with an empty response when Claude reached for a tool mid-run.
    • A successful run whose content merely mentions an auth error is no longer misreported as an expired Claude session — auth failures are flagged only when the run actually fails.
    • Memophant now uses its bundled memory server instead of rebuilding it on every launch: faster startup, no Xcode/Swift toolchain required, and adding a project can't dead-end in a “binary missing” loop.

  8. 1.2

    Memory Review becomes a full workspace, and how your notes anchor to code is finally your call.

    New

    • Memory Review is a full workspace: see each flagged note's actual code diff, edit or delete it in place, and resolve whole groups at once — filtered by what each note needs (re-verify, ground, fix a broken link, retire).
    • A running “N resolved · M remain” shows exactly what you cleared as you work through the inbox.
    • Grounding is on your terms — choose Auto-ground all for fast best-guess coverage, or review each note to declare the anchors yourself. Nothing gets anchored behind your back.

    Improved

    • Fewer false alarms: a comment tweak or a formatting pass no longer flags every memory on that file — only real, semantic code changes do.
    • A note that merely mentions a file in passing is no longer anchored to it, so it can't fire phantom drift.
    • Honest health: durable decisions and conventions are no longer nagged to anchor to code they don't track, and stray guessed anchors from older versions are cleaned up automatically.

    Fixed

    • Consolidation's “Re-run…” no longer loops forever on a note that can't be audited — it's surfaced for manual review after a few tries.
    • Background git operations can no longer silently strand a worker thread under heavy concurrent load.
    • Hardened task-board concurrency.

  9. 1.1

    Trustworthy memory: your project's memory now knows when it's stale, heals itself, and keeps a consistent, searchable shape — with every AI-proposed change behind your review.

    New

    • Freshness is grounded in code drift, not the calendar — notes anchor to the files they document, and a note stays green until that code actually changes.
    • A consolidation engine you can trust: scope-bounded agentic audits run in small batches, investigate only evidence-backed drift, and resume exactly where an interrupted pass left off.
    • An adversarial verifier independently re-checks every proposed change against your real code — refute-by-default, with the reason shown — before it reaches your review.
    • Significant decisions get promoted into a curated History on the topic's wiki page, with links both ways and automatic re-promotion when the source note drifts.
    • The code tier goes multi-language: TypeScript, JavaScript, Python, Go, and Rust symbol extraction, with per-page drift dots.

    Improved

    • The agent surface grew substantially — memory health, review, retire, and documents/templates access.
    • Writes are fast and safe: single-note indexing ends the full-corpus re-embed hangs, and titles with slashes or punctuation can no longer create phantom folders.
    • The MCP server now ships prebuilt inside the app — no Swift toolchain needed — and tells agents when a rebuilt binary needs a session restart.
    • Semantic search now reads prose notes, not just structured bullets, and notes render their full body and real titles.
    • The task board works from any git worktree, with a priority filter and safer concurrent edits.

    Fixed

    • Retired notes stop nagging, and orphaned notes whose source code was deleted get their own honest signal instead of a false green.
    • Sidebar width persists across launches, and the download is a notarized, stapled DMG.

  10. 1.0.1

    Launch-week fixes and polish.

    Improved

    • The download link is now stable across releases — no per-release URL to update.
    • Agents are guided to file every memory note under one of the six canonical folders, with a gentle in-place correction when a note lands in the wrong place.

    Fixed

    • The secret scanner no longer false-positives on ordinary kebab-case slugs that happen to start with “sk-”.
    • Long area names in the dashboard's corpus footprint wrap and truncate cleanly instead of overflowing, and root-level notes bucket under “Ungrouped.”
    • Code search joins multi-word queries correctly across all three search surfaces.

  11. 1.0

    Memophant 1.0 — repo-resident AI project memory, publicly launched.

    New

    • Give your AI agents a project memory that lives in your repo as plain files — versioned in git, readable by you, and shared across every editor through MCP.
    • Layered memory across tiers — decisions, conventions, architecture, operations, and more — plus a wiki and code index, all searchable from a native macOS command center.
    • Drift detection flags notes as stale when the code they describe moves on, so an agent never acts on out-of-date facts.
    • Automatic updates via Sparkle, a notarized Universal build, and a one-time Heirloom license.

Full commit history and signed builds live on GitHub.

Memory that keeps getting sharper.

Every release makes your project's memory more trustworthy — and updates arrive automatically.

Download for macOSSee the system