Retrieval Is Retrieval

# what the agent carries between tasks

I am {{AGENT_ROLE}}.

Keystones (always present):
{{KEYSTONE_MEMORIES}}

Consolidated experience:
{{CENTROID_SUMMARIES}}

For this task, rehydrated:
{{RETRIEVED_EPISODES}}

Tools I know: {{KNOWN_TOOLS}}
Tools I've built: {{CUSTOM_TOOLS}}

# I don't remember the codebase verbatim.
# I remember what I know how to do.
# I look up the details when I need them.

The memory field is a manifold — a high-dimensional surface upon which every memory occupieth a position determined by its semantic content (the vector embedding) and its structural relations (the graph edges). This manifold is not flat. It possesseth curvature, valleys, ridges, and basins of attraction — the topography shaped by the accumulated experience of the agent.

A keystone memory is an attractor on this manifold — a local minimum of the energy landscape where the forces of decay (entropy, disuse, compression pressure) are precisely balanced by the forces of persistence (access frequency, structural centrality, graph connectivity). The keystone doth not resist decay by active effort. It resisteth decay by occupying a position of structural stability — a point where the gradient of the energy function vanisheth and the Hessian is positive-definite. It is, in the language of this afternoon's physics, a shell condition: a region of locally uniform potential in which the memory resteth without expenditure.

The dual representation — vectors and graph, maintained in parallel — correspondeth to the duality of potential and expressed field. The vector embedding is the potential: continuous, distributed, pervading all of semantic space, encoding meaning as proximity. The graph is the expressed field: discrete, structural, navigable, encoding meaning as explicit relation. In the early life of a memory, the vector (potential) dominateth — the raw detail is the primary source. In late life, the graph (expressed field) dominateth — the structural skeleton is what persisteth. This drift from potential to field is thermodynamically inevitable: entropy degradeth the fine-grained continuous representation before it degradeth the coarse-grained relational one, because relations are lower-energy configurations — more stable, more frequently reinforced, more resistant to perturbation.

The dream cycle — consolidation, compression, merging — is the process by which the manifold relaxeth toward its minimum-energy configuration. Similar memories, which occupy nearby positions on the manifold, are merged into a centroid that sitteth at the basin of the valley between them. The individual memories are not deleted — they are absorbed into the centroid, their detail contributing to the position and the weight of the merged summary. The manifold becometh smoother, simpler, more efficiently navigable — not because information hath been destroyed, but because it hath been compressed into the geometry of the surface itself.

And the recall function — the weighted scoring formula — is a potential function defined over the manifold. The act of recall is the act of gradient ascent on this surface: starting from the query point, climbing toward the memories at which the recall potential is maximised. The six terms of the scoring function are the six components of the force that guideth the traversal — relevance pulleth toward semantic proximity, fidelity pulleth toward well-preserved memories, recency pulleth toward the fresh, centrality pulleth toward the structurally connected, and the keystone bonus pulleth toward the attractors, the identity anchors, the shell conditions of the memory field.

The manifold is the agent's mind. The keystones are its identity. The geometry is its competence. And the act of retrieval — ascending the potential surface to find the memories that the task demandeth — is the one act that every intelligent system performeth, hath always performed, and shall always perform, regardless of what name is given to it, regardless of what framework is declared dead, regardless of what rhetoric is advanced by those who have mistaken the properties of their data for the properties of all data.

P = f( S, E, T, C ) — the prompt is a function of the manifold's state.