Claim Boundary
Strongest Independent System Claim
A system for coordinating multiple LLM agents within a mission, comprising:
a mission-scoped shared vector collection created on mission start and destroyed on mission completion
an injection mechanism that converts agent task outputs into high-dimensional embedding vectors with metadata (agent ID, strength, timestamp, content hash) and stores them in the shared collection
a query mechanism that retrieves patterns by semantic similarity and ranks them by a resonance score combining squared cosine similarity, exponential temporal decay from last access, and access-count-based reinforcement
agent-callable platform tools enabling agents to query, inject, and measure the shared field during task execution
Why this is the strongest system claim: It captures the full architecture — the shared field, the scoring function, the lifecycle binding, and the agent interface — as a single coordinated system. Each element alone exists in prior art. The combination applied to multi-agent LLM coordination does not appear in any known system.
Strongest Independent Method Claim
A method for reducing information loss in sequential multi-agent LLM pipelines, comprising:
creating a shared embedding space at mission start
automatically injecting each agent's completed task output into the shared space
enabling any downstream agent to query the shared space by semantic meaning, bypassing intermediate agents
ranking results by resonance (squared cosine similarity x temporally-decayed strength x access reinforcement)
reinforcing patterns that are retrieved (Hebbian access boost) and co-retrieved (co-access strength bonus)
destroying the shared space when the mission terminates
Why this is the strongest method claim: It focuses on the process — what happens and in what order — rather than the system components. This is harder to design around because competitors would need to change the workflow, not just swap out a component.
Dependent Claim Ideas
D1
System
Content deduplication via SHA-256 hash with reinforce-on-collision
D2
System
Field stability metric (weighted mean strength + organization score)
D3
System
Boundary permeability coefficient controlling injection strength
D4
System
Over-fetch factor (3x top_k) compensating for archival filtering
D5
Method
Seeding the shared space with the mission goal as the first pattern
D6
Method
Archival threshold filtering at query time (non-destructive)
D7
System
Abstract interface (SharedContextPort) enabling backend substitution for A/B comparison
D8
Method
Co-access bonus mutating stored strength (not just query-time scoring)
What We Are NOT Claiming
We do not claim invention of:
Vector embeddings or cosine similarity
Temporal decay or exponential decay functions
Hebbian learning as a concept
Blackboard architectures or shared workspaces in general
Semantic search or approximate nearest neighbor retrieval
Vector databases (Qdrant, Pinecone, etc.)
Multi-agent systems or agent orchestration in general
The concept of shared memory for software agents
We do not claim:
Universal superiority over all multi-agent coordination approaches
That message passing is obsolete or broken in all contexts
Academic proof of effectiveness across diverse workloads
Exhaustive prior-art search (counsel should conduct formal search)
That any single element of this system is novel in isolation
Claim Scope Assessment
Broadest defensible scope: The specific combination of mission-scoped shared vector field + resonance scoring (cos^2 x decay x reinforcement) + mission lifecycle binding + agent-callable tools, applied to multi-agent LLM coordination.
Narrowest fallback scope: The resonance scoring formula (cos^2 x temporally-decayed strength x access boost with co-access bonding) applied to a shared multi-agent knowledge field, with content-hash deduplication using reinforce-on-collision.
Counsel should advise on: Whether the broadest scope survives a formal prior-art search, and which dependent claims add the most defensive value.
Last updated