a living memory tapestry woven for AI agents
Voidloom proposes memory connections nobody asked it to find
We use this standalone agent memory across sessions; later recall can surface its proposals, including a forgotten precedent.
Voidloom pairs distant and dissonant memories overnight
It records the proposed connections as reflections at low confidence, which ordinary recall can return later.
You rebuild a solution before finding its precedent in notes from an older project. After nightly consolidation, Voidloom dreams by pairing memories across contexts. A connection that conflicts with existing memories stays tentative and fades unless new evidence reinforces it. Recall returns dreams in a clearly labeled channel beside the memory results, from the same call. Each channel is ranked separately; a dream never takes a recorded memory’s place.
Voidloom keeps agent memory in a single SQLite database file on your machine.
As of , Voidloom’s background thinking runs on the hosted or local model chosen at Voidloom’s own setup, if any. With a local model, research into gaps in what Voidloom knows is skipped because that research needs the web.
Like that database file, the rest of Voidloom’s structure is familiar: a vector index supports recall by meaning, a full-text index supports exact words, and a knowledge graph holds typed links between memories. Voidloom can run standalone without Voidstead (a home for AI agents, being built for your own machine). Within that home, Voidloom comes built in as its memory; its installation inside Voidstead is being completed.
Voidloom explains why it recalled a memory
Recall combines meaning and exact words.
For each recalled item, Voidloom can explain the match factor by factor, so you can inspect why the memory returned.
Voidloom leaves memory disagreements for you to resolve
A proposed revision needs your approval.
Voidloom identifies disagreements between memories about the same subject, without inventing conflicts between unrelated memories. A retired belief keeps its history.
Voidloom proposes changes to its own parts for your approval
A monthly evaluation identifies parts to improve or retire.
Day to day, Voidloom grows on its own: it stores what agents record, consolidates memories, prunes links that fade, and tests its recall. With a hosted or local thinking model chosen at setup, Voidloom dreams reflections held at low confidence; with none, it produces no reflections. Research into Voidloom’s knowledge gaps requires a hosted model. Anything learned from the web stays marked as outside knowledge.
Voidloom gives agents context when sessions would otherwise start from zero
We have used Voidloom every day in our own work for months.
Memory grows, recalls, and proposes connections
This map shows how Voidloom keeps and recalls memories, proposes connections, and tests its recall. We drew it for this page; it is not a screen from Voidloom. Every star marks a working capability.
- Stars: working capabilities
The Bare Thread
Memory persists and gives agents session context.
The First Pattern
Recall finds memories; events become lessons overnight.
The Living Tapestry
Sources stay traceable; important older memories resurface.
The Loom That Re-Weaves Itself
Connections proposed; dreamed contradictions stay tentative.
The Loom That Weaves Looms
Disagreements, approvals, recall reasons, and self-checks.
The Bare Thread
Memory persists and gives agents session context.
The First Pattern
Recall finds memories; events become lessons overnight.
The Living Tapestry
Sources stay traceable; important older memories resurface.
The Loom That Re-Weaves Itself
Connections proposed; dreamed contradictions stay tentative.
The Loom That Weaves Looms
Disagreements, approvals, recall reasons, and self-checks.
Memory proposes connections overnight that an agent can encounter during later recall.
Memory carries over What agents record remains in memory across sessions.
Recall on bare machines Keyword recall works without a model or a network.
Context when sessions begin Voidloom gives an agent context at the start of each session.
More than one coding agent Its memory works with more than one coding agent.
Meaning and exact words Recall combines matches by meaning and by exact words.
Daily events become lessons Each night, Voidloom distills the day’s events into lessons.
Memories used together strengthen It strengthens memories used together, lets others recede, and prunes fading links.
Retired beliefs keep their history A retired belief retains the history of what was believed.
Lessons keep their sources A lesson records its source observations, so its origin can be traced.
Important older memories resurface Each week, important older memories resurface so recent notes do not bury them.
Outside knowledge stays labeled What Voidloom learns from the web stays permanently marked as outside knowledge.
Dreams return separately Dreams return in a labeled channel beside memories; each channel is ranked separately.
Connections nobody asked for Voidloom deliberately pairs distant and dissonant memories to propose connections.
Dreamed conflicts stay tentative A dreamed contradiction stays weak and fades unless new evidence reinforces it.
Precedents can return later Later recall can bring back a forgotten precedent from an overnight reflection.
Improvements proposed for approval Monthly evaluations propose improvements or retirements of its parts for your approval.
Thinking model choice You choose Voidloom’s background thinking model at setup: hosted, local, or none for now.
Genuine disagreements are reported It reports memories that disagree about the same subject and proposes a revision.
Belief and component changes wait Belief revisions and changes to its own parts wait for your approval.
Recall explains its matches Voidloom shows why a memory was recalled, factor by factor.
Tests its own recall It tests its recall and records weak recall as a gap to research, never a stored guess.
Notices parts with no output It distinguishes parts that ran with output from those that ran but produced nothing.
Read the map
The Bare Thread
Memory persists and gives agents session context.
Recorded memories persist, and an agent receives context when a session begins.
- Memory carries over: What agents record remains in memory across sessions.
- Recall on bare machines: Keyword recall works without a model or a network.
- Context when sessions begin: Voidloom gives an agent context at the start of each session.
- More than one coding agent: Its memory works with more than one coding agent.
The First Pattern
Recall finds memories; events become lessons overnight.
Recall uses meaning and exact words; nightly consolidation turns events into lessons.
- Meaning and exact words: Recall combines matches by meaning and by exact words.
- Daily events become lessons: Each night, Voidloom distills the day’s events into lessons.
- Memories used together strengthen: It strengthens memories used together, lets others recede, and prunes fading links.
- Retired beliefs keep their history: A retired belief retains the history of what was believed.
The Living Tapestry
Sources stay traceable; important older memories resurface.
Lessons remain traceable to their sources, and important older memories return each week.
- Lessons keep their sources: A lesson records its source observations, so its origin can be traced.
- Important older memories resurface: Each week, important older memories resurface so recent notes do not bury them.
- Outside knowledge stays labeled: What Voidloom learns from the web stays permanently marked as outside knowledge.
- Dreams return separately: Dreams return in a labeled channel beside memories; each channel is ranked separately.
The Loom That Re-Weaves Itself
Connections proposed; dreamed contradictions stay tentative.
An overnight connection can bring a forgotten precedent into later recall.
- Connections nobody asked for: Voidloom deliberately pairs distant and dissonant memories to propose connections.
- Dreamed conflicts stay tentative: A dreamed contradiction stays weak and fades unless new evidence reinforces it.
- Precedents can return later: Later recall can bring back a forgotten precedent from an overnight reflection.
- Improvements proposed for approval: Monthly evaluations propose improvements or retirements of its parts for your approval.
- Thinking model choice: You choose Voidloom’s background thinking model at setup: hosted, local, or none for now.
The Loom That Weaves Looms
Disagreements, approvals, recall reasons, and self-checks.
You can inspect recall and decide on belief revisions and changes to Voidloom’s parts.
- Genuine disagreements are reported: It reports memories that disagree about the same subject and proposes a revision.
- Belief and component changes wait: Belief revisions and changes to its own parts wait for your approval.
- Recall explains its matches: Voidloom shows why a memory was recalled, factor by factor.
- Tests its own recall: It tests its recall and records weak recall as a gap to research, never a stored guess.
- Notices parts with no output: It distinguishes parts that ran with output from those that ran but produced nothing.
Questions about Voidloom
What is Voidloom?
Voidloom is memory for AI agents that proposes unrequested connections for later recall. It stands on its own and supplies context when a session begins. Recall combines meaning with exact wording and can later return an overnight proposal, including a forgotten precedent.
Is Voidloom’s dreaming just a summary of related memories?
No. Voidloom seeks connections across memories that seem unrelated or disagree. A dreamed contradiction remains a weak proposal until further evidence strengthens it; without that evidence, it fades.
Does Voidloom explain how the model reasoned?
No. Voidloom explains why its own recall returned a memory, factor by factor, not how a model reasoned.
Does the memory silently choose which belief wins?
No. Voidloom proposes a revision when memories genuinely disagree, and you decide whether to accept it. That decision can revise a belief or retire it.
Does Voidloom need approval to grow?
No. Its routine growth proceeds on its own. Approval is required to revise or retire a belief after a disagreement, or to improve or retire a component. Voidloom never rewrites its own code.
Does basic keyword recall need a model to be installed?
No. Voidloom can retrieve memories by keyword on a bare machine, with no model or network connection. Recall by meaning also runs on your own machine without outside services once its model is installed.
Can a lesson in Voidloom be traced to where it came from?
Yes. When Voidloom distills observations into a lesson, it records the observations behind that lesson. The lesson can then be traced back to those observations.
Overnight, Voidloom draws threads between memories from different contexts, so a forgotten precedent can surface the next time an agent recalls.
In private use. Finishing touches in development; not yet released.
What it does not do ·
- Its integration into Voidstead, a home for AI agents, being built for your own machine, is not finished yet.
- With a hosted model chosen at setup, seven background steps, including consolidation, dreaming, and monthly evaluation, send memory content to its provider under your own account. A local model receives that content on your machine. If a chosen model is unavailable, consolidation and monthly evaluation use simpler rules; dreaming and gap research produce nothing. With none chosen, simple nightly routines run without background thinking or sending memory content out.
- With a local model, Voidloom skips research into knowledge gaps and logs the omission. With a hosted model, research sends a topic name and the kind of gap to that model, which searches the web.
- Installation does not proceed if the chosen model fails the setup test: a question the model must answer correctly. Voidloom explains why installation stopped.
- Recall does not yet keep a retired belief from surfacing, and no test covers that behavior yet.
Thank you. We will email you when Voidloom is released.
Thank you for sharing.