Avid memory stack

The Avid memory stack divides one agent’s memory into four layers: working context, episodic, semantic, and personal.¹ A dream cycle promotes recurring episodic lessons into the semantic layer, and personal preferences stay out of general lessons.¹

Overview#

Forgetting drops or revises facts that contradict newer evidence.² A semantic or procedural edit is kept only when a paired check shows that it helped.³⁴

Mechanism#

1. Preferences are injected first, then semantic lessons, then episodes, then only the skills that match.¹
2. High-salience episodes are promoted into durable lessons, and material is archived rather than silently deleted.¹
3. Contradictions are pruned so the remaining layers stay trustworthy.²
4. Procedures are kept as their own graph, which biases the next step without dictating it.⁴
5. Memory is kept in files under version control and synced only on purpose.⁵

Applications#

The stack applies when flat memory is collapsing and one agent needs layers, promotion, and pruning.

Limitations#

The stack is a single agent’s memory, not a team’s write gate (see Shared brain). A vector store is not added before a recall problem has been measured. Not every episode belongs in always-on context.

Worked example#

EvoOntology applies the keep rule to a semantic layer for data agents.³

1. A dedicated builder agent constructs the semantic layer (an ontology) instead of a person pasting one into the prompt.
2. The layer is served as a queryable server with a schema layer, a content layer, and a tool layer, and the data agent queries it at runtime.
3. A self-evolution loop proposes small, typed edits to the layer, guided by attribution.
4. An edit is accepted only after a paired evaluation on the same backbone model shows that it helped; otherwise it is dropped.

See also#

• Skillify authoring – the skills on-ramp⁶
• GBrain – a product pointer
• Shared brain – a search-first team layer
• Thin harness, fat skills – the thin harness
• Karpathy LLM wiki foundation – compiling from raw sources

Further reading#

• https://x.com/omarsar0/status/2097755424007373270
• https://github.com/codejunkie99/agentic-stack
• https://github.com/codejunkie99/brain
• https://x.com/av1dlive/status/2052063154423898603

References#

Footnotes#

1. https://x.com/av1dlive/status/2044453102703841645 ↩ ↩² ↩³ ↩⁴

2. https://x.com/the0xbt/status/2102166829007323180 ↩ ↩²

3. https://arxiv.org/abs/2609.15779 ↩ ↩²

4. https://arxiv.org/abs/2609.09153 ↩ ↩²

5. https://x.com/av1dlive/status/2048800729985634397 ↩

6. https://x.com/av1dlive/status/2053778835100119288 ↩