Claims

Independent Claims

Claim 1 — Multi-Agent Shared Semantic Field Coordination

A computer-implemented method for coordinating a plurality of large language model (LLM) agents executing tasks within a mission, the method comprising:

(a) creating, responsive to initiation of a mission comprising a plurality of agent tasks, a shared vector embedding collection in a vector database, the collection being scoped to said mission;

(b) for each agent task that completes during said mission, automatically injecting the task output into the shared collection as a vector point comprising a high-dimensional semantic embedding of the output content, an agent identifier, a strength value, a timestamp, and a content hash;

(c) during execution of a subsequent agent task, receiving from an executing agent a natural language query directed to the shared collection;

(d) computing, for each candidate vector point in the collection, a resonance score based on (i) the squared cosine similarity between the query embedding and the candidate embedding, (ii) a temporal decay factor computed from the elapsed time since the candidate was last accessed, and (iii) an access-based reinforcement factor derived from the candidate's retrieval count;

(e) returning to the executing agent the top-k candidate vector points ranked by resonance score; and

(f) destroying the shared collection responsive to the mission reaching a terminal state.

Claim 2 — Resonance-Based Retrieval Scoring

A computer-implemented method for ranking query results in a shared multi-agent knowledge field, comprising:

(a) receiving a query embedding from an agent;

(b) retrieving candidate patterns from a vector database by approximate nearest neighbor search;

(c) for each candidate pattern, computing a resonance score according to the formula:

R = cos^2(theta) x S_0 x e^(-lambda x t) x min(1 + alpha x n, C)

where:

• theta is the angle between the query embedding and the candidate embedding,

• S_0 is the stored strength of the candidate pattern,

• lambda is a configurable decay rate,

• t is the elapsed time in hours since the candidate was last accessed,

• alpha is a per-access reinforcement increment,

• n is the candidate's access count, and

• C is a reinforcement cap;

(d) filtering candidates whose temporally-decayed strength falls below a configurable archival threshold; and

(e) returning the remaining candidates ranked by resonance score.

Claim 3 — Hebbian Co-Access Reinforcement in Agent Knowledge Fields

A computer-implemented method for strengthening associative relationships between knowledge patterns in a shared agent field, comprising:

(a) executing a semantic query against the shared field on behalf of an agent;

(b) retrieving a result set of k patterns;

(c) for each pattern in the result set, incrementing an access count and updating a last-accessed timestamp;

(d) for each pattern in the result set, computing a co-access strength bonus based on the number of co-retrieved patterns:

S_new = min(S_current x (1 + beta x (k - 1)), S_current x C)

where beta is a co-access bonus coefficient and C is a reinforcement cap; and

(e) persisting the updated strength value to the vector database, whereby patterns that are frequently retrieved together in response to agent queries become mutually strengthened.

Dependent Claims

Claim 4

The method of Claim 1, further comprising, prior to step (b), seeding the shared collection with the mission goal as an initial pattern at a predetermined strength value, such that agent queries can discover the mission objective by semantic similarity.

Claim 5

The method of Claim 1, wherein step (b) further comprises:

• computing a SHA-256 hash of the content to be injected;

• querying the collection for an existing pattern with a matching content hash;

• if a matching pattern exists, reinforcing the existing pattern by incrementing its access count and updating its last-accessed timestamp instead of creating a new vector point.

Claim 6

The method of Claim 1, wherein the shared collection is implemented behind an abstract interface comprising create_context, inject, query, and destroy_context methods, enabling substitution of the vector field implementation with an alternative coordination backend without modification to mission orchestration code.

Claim 7

The method of Claim 2, wherein the retrieval step (b) fetches a multiple of the requested top-k results from the vector database (over-fetching), applies the archival filtering of step (d), and returns the final top-k from the remaining candidates, compensating for candidates eliminated by the archival threshold.

Claim 8

The method of Claim 1, further comprising computing a field stability metric:

stability = avg_strength x w_1 + organization x w_2

where avg_strength is the mean temporally-decayed strength of all patterns, organization is computed as max(0, 1 - (standard_deviation / mean)) of pattern strengths, and w_1 and w_2 are configurable weights, said metric indicating the degree to which agents are building shared understanding versus the field fragmenting.

Claim 9

The method of Claim 1, wherein step (b) scales the initial strength of each injected pattern by a configurable boundary permeability coefficient between 0.0 and 1.0, controlling the rate at which new knowledge enters the field.

Claim 10

The method of Claim 1, wherein the plurality of agents interact with the shared collection through agent-callable platform tools comprising:

• a query tool that accepts a natural language query and returns resonance-ranked patterns;

• an inject tool that accepts a semantic key and value and creates a new pattern with deduplication; and

• a stability tool that returns the field's current convergence metrics;

said tools being registered through a platform tool discovery pipeline and available to any agent assigned to a mission with an active field.