Engram Alpha
A project memory for AI coding assistants: why things were decided and what went wrong, kept as a graph you can see and edit.
by techtheist
In short
While you code, your AI assistant writes down the decisions, gotchas, problems and fixes worth keeping, as typed notes joined by links like “replaces”, “because” and “conflicts with”. Each new session opens with a short brief of the project's standing knowledge and open work, and the assistant can search the rest. Notes that look alike are flagged for you or the assistant to judge, and you can read, edit and review the whole graph in a pane in your browser or IDE.
Who it’s for
Developers who use AI coding assistants (Claude Code, Codex, Gemini CLI and others) and want the project's reasoning to outlast each session, on their own machine.
What it remembers
A project's reasoning: decisions and their reasons, principles, cautions, problems and how they were solved, insights, and open to-dos. Not code structure, and not you as a person.
- Memory lives
- A file in your repo, on your machine
- API keys
- None
- Works with
- Claude Code, Codex, Gemini CLI, OpenCode, Kilo, Google Antigravity, Windsurf, Devin CLI, Bob (IDE and Shell), JetBrains IDEs, VS Code
- License
- Open sourceMIT
- Price
- Free
Under the hood
- Who writes memories
- The AI assistant you're working with, during the session, guided by a skill with three levels (relaxed, normal, aggressive). You can write and edit notes by hand in the pane.
- Where they're kept
- One database file inside your repo (TepinDB, or SQLite), optionally encrypted. Exportable as JSON.
- How they're found
- Keywords and meaning (local embeddings), re-ranked by a small local model and weighted by trust. Each hit comes with its linked notes, conflicts and replacements first.
- Extra AI calls
- No AI calls of its own. Embedding, re-ranking and contradiction checks run on small local models; the assistant's reads and writes are ordinary tool calls in your session.
How it works
- 1Session startsThe assistant is handed a brief of standing decisions, cautions and open work.
- 2It searchesBefore real work, it looks up what's known, linked notes included.
- 3It writesDecisions and gotchas are saved quietly, each checked against nearby notes.
- 4It judgesLook-alikes and contradictions are settled as conflict, replaces or fine.
- 5You reviewIn the graph pane, approve, edit, pin or delete what it kept.
What stands out
- The memory is a graph you can see and edit, inside JetBrains IDEs, VS Code or a browser.
- It checks a claim against what the project already decided, on a small local model, and lists the notes that support or contradict it.
- Replaced knowledge is kept as history, and deliberately removed knowledge leaves a tombstone, so it isn't learned again.
- Its retrieval settings come from a benchmark that ships with it and can be re-run.
What it doesn’t do
- It doesn't write memories itself: the AI assistant you're working with decides what to save, and you can review it.
- Being found in a search doesn't make a memory stronger. Only an edit, a confirmation or your approval does.
- Nothing merges, replaces or corrects a memory on its own: the local models point out look-alikes and contradictions, and the assistant or you decide.
- No emotional weighting, no dated reminders, and no memory about you as a person.
- It's for one person on one machine; there's no team sync.
Measured against human memory
3 built (2 in depth), 6 partly, 2 not of the Field Guide’s 11 mechanisms
Encoding
Salience
Decides what's worth keeping, and how much, as a memory is written.
The assistant keeps or skips by written rules (three capture levels) and labels each note with a type and how long it should matter. There's no importance score. source
Emotion
Gives memories written with feeling more weight.
No feelings are recorded or weighed.
Storage
Forgetting
Lets unused memories fade over time on its own.
How fast a note fades depends on its kind: short-lived notes over about a month, episodes over about six months, decisions and principles never with time alone. Long-faded notes nobody approved are archived, not deleted. source
Interference
When a new memory closely matches an old one, notices and lets the newer one win.
Each write is checked against nearby notes: a near-copy is folded into the existing note, and look-alike pairs wait for you or the assistant to judge them. The newer one never wins on its own; when it's judged to replace the older, the older is archived behind a “replaces” link. source
Retrieval
Retrieval
Remembering something makes that memory stronger.
It records when a note was last found, but being found never makes it stronger; only an edit, a confirmation or your approval does. source
Association
Links memories, so recalling one brings its neighbours.
Notes are joined by typed links (because, answers, replaces, conflicts with, and others). Every search hit brings its linked notes, and the assistant can walk the graph further. source
Prospective
Brings a memory back at the right time without being asked.
When the assistant opens or edits a file, notes that point at that file come back unasked. Each session also opens with the open work, and a note tagged “handoff” leads the next brief. source
Transformation
Consolidation
Reworks memories between sessions: strengthening, merging, reorganising.
A background scan every six hours, and sweeps you start, find look-alikes, hidden contradictions and faded notes, and queue them for review. Nothing is merged without a judgement, and every change is in the audit log. source
Reconsolidation
Corrects a memory when it's contradicted, and keeps what it used to say.
A correction archives the old note and keeps every version on a timeline. Decisions lose trust only to a judged contradiction, a pinned note not even then, and you can search for what was known at a given time or version. source
Gist
Turns many specific moments into general knowledge.
The assistant writes insights and principles when it notices them, but no step draws general lessons from many notes. source
Schemas
Forms beliefs about people and projects, and weighs new facts against them.
Every new note is checked against nearby decisions and principles: does what's known support or contradict it. Each note carries a trust value. There's no standing summary of the project, and no picture of people. source
Not every system is trying to be a human memory, so a hollow dot isn’t a flaw. It just says what the system doesn’t do. What the mechanisms are.
Engram Alpha’s makers haven’t shared what they’re building next, so only today shows here. Makers can add their roadmap the same way they correct an entry.
Compare Engram Alpha with the others, side by sideQuestions
- Who is Engram Alpha for?
Developers who use AI coding assistants (Claude Code, Codex, Gemini CLI and others) and want the project's reasoning to outlast each session, on their own machine.
- Where does Engram Alpha keep its memories?
A file in your repo, on your machine. One database file inside your repo (TepinDB, or SQLite), optionally encrypted. Exportable as JSON.
- Does Engram Alpha forget?
Forgetting: built in depth. How fast a note fades depends on its kind: short-lived notes over about a month, episodes over about six months, decisions and principles never with time alone. Long-faded notes nobody approved are archived, not deleted. source
- Does Engram Alpha correct a memory when it turns out to be wrong?
Reconsolidation: built in depth. A correction archives the old note and keeps every version on a timeline. Decisions lose trust only to a judged contradiction, a pinned note not even then, and you can search for what was known at a given time or version. source
- Is Engram Alpha open source?
Yes, it's open source. License: MIT. Code: github.com/techtheist/engram.
- Does Engram Alpha work with Claude Code?
Yes. Works with: Claude Code, Codex, Gemini CLI, OpenCode, Kilo, Google Antigravity, Windsurf, Devin CLI, Bob (IDE and Shell), JetBrains IDEs, VS Code.
- Does Engram Alpha need an API key?
API keys: None.
- How much does Engram Alpha cost?
Price: Free.