# US 2026/0310691 A1 · Dream Memory for Providing Stored Playbooks to LLMs

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2026/0310691 A1
- Document ID: US20260310691A1
- Status: Published application
- Kind: A1
- Application number: 19/659,292
- Filed: 2026-04-27
- Published: 2026-10-08
- Earliest claimed: 2023-05-04
- Assignees: Vijay Madisetti
- Inventors: Vijay Madisetti; Arshdeep Bahga
- Web page: https://patentgrep.com/patent/US20260310691A1

## CPC

- G06F 16/3329: Natural language query formulation
- G06F 40/284: Lexical analysis, e.g. tokenisation or collocates
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

Systems and methods for augmenting reasoning capabilities of LLM agents through memory-augmented reasoning enhanced by synthetic reasoning patterns. A problem distiller module extracts a distilled problem representation from an input problem. A retrieval module queries a persistent memory store to retrieve relevant reasoning templates based on similarity to the distilled problem representation. A template instantiation module adapts retrieved templates to the problem context by binding abstract strategy descriptions and generalized procedural steps to the distilled problem representation. A dream generator module generates synthetic reasoning patterns through parameterized prompt templates, each encoding a structured problem-solving strategy belonging to functionally distinct reasoning pattern categories. A prompt assembly module assembles a structured prompt processed by a reasoning engine comprising the LLM agent to generate a solution. A quality evaluator module evaluates the solution to produce a quality score. A memory updater module updates the persistent memory store based on the quality score.

## Claims

### Claim 1

A computer-implemented method executable by one or more processors for augmenting the reasoning capabilities of a large language model (LLM) agent, the method comprising:
receiving, by a problem distiller module, an input problem;
extracting, by the problem distiller module, a distilled problem representation from the input problem;
querying, by a retrieval module, a persistent memory store containing a plurality of stored reasoning templates using the distilled problem representation;
retrieving, by the retrieval module, one or more relevant reasoning templates based on a similarity to the distilled problem representation;
adapting, by a template instantiation module, each relevant reasoning template of the one or more relevant reasoning templates to a specific context of the input problem by binding one or more abstract strategy descriptions and one or more generalized procedural steps to the distilled problem representation, thereby producing one or more instantiated templates;
generating, by a dream generator module, one or more synthetic reasoning patterns by processing the distilled problem representation through one or more parameterized prompt templates, each synthetic reasoning pattern being configured to encode a structured problem-solving strategy belonging to one of a plurality of functionally distinct reasoning pattern categories;
assembling, by a prompt assembly module, a structured prompt;
processing, by a reasoning engine comprising the LLM agent, the assembled structured prompt to generate a solution to the input problem;
evaluating, by a quality evaluator module, the solution to produce a quality score; and
updating, by a memory updater module, the persistent memory store from a current memory state to an updated memory state based on the quality score, the distilled problem representation, the one or more instantiated templates, the generated synthetic reasoning patterns, and the solution, whereby stored reasoning templates that contributed to a solution having a quality score above a threshold quality score value are promoted in retrieval ranking and stored reasoning templates associated with an incorrect solution are demoted, thereby enabling progressive improvement of the reasoning capabilities of the LLM agent over successive problem-solving interactions.

### Claim 2 (depends on 1)

The method of claim 1, wherein the plurality of functionally distinct reasoning pattern categories comprises at least one of:
strategy dreams configured to encode domain-specific heuristics that reduce the solution search space;
meta pattern dreams configured to encode abstract, domain-transcendent reasoning patterns comprising higher-order problem-solving approaches;
error avoiding dreams configured to encode common pitfalls and systematic verification guidelines;
hard problem dreams configured to encode specialized techniques for problems exceeding a predetermined difficulty threshold;
adaptive dreams configured to be dynamically generated based on specific characteristics of the current problem instance; and
negative examples that comprise an incorrect solution paired with an explanation identifying the errors contained therein.

### Claim 3 (depends on 2)

The method of claim 2, wherein:
the dream generator module is configured to generate each synthetic reasoning pattern of the one or more synthetic reasoning patterns by supplying a problem domain descriptor and a problem description as parameterized inputs to a domain-agnostic prompt template corresponding to a reasoning pattern category associated with the synthetic reasoning pattern; and
each prompt template is configured to produce a structured output comprising at least one of:
a name field;
a description field;
an application procedure field;
an illustrative example field; and
a conditions field.

### Claim 4 (depends on 2)

The method of claim 2, wherein:
the adaptive dreams are generated at inference time in response to specific characteristics detected in the distilled problem representation; and
the strategy dreams, the meta pattern dreams, the error avoiding dreams, the hard problem dreams, and the negative examples are generated prior to deployment and stored persistently in the persistent memory store.

### Claim 5 (depends on 1)

The method of claim 1, wherein the distilled problem representation comprises:
one or more identified features of the input problem;
one or more constraints of the input problem; and
one or more structural characteristics of the input problem.

### Claim 6 (depends on 1)

The method of claim 1, wherein the structured prompt comprises:
a rules block specifying one or more domain constraints;
a problem block including the distilled problem representation; and
a dreams block including the instantiated templates and the generated synthetic reasoning patterns.

### Claim 7 (depends on 1)

The method of claim 1, wherein extracting the distilled problem representation comprises:
identifying, by a feature extraction unit, one or more of a numerical feature, a structural feature, or a semantic feature of the input problem;
enumerating, by a constraint identifier, at least one of explicit and implicit constraints governing valid solutions; and
characterizing, by a structure analyzer, a problem type and a complexity level associated with the input problem.

### Claim 8 (depends on 1)

The method of claim 1, wherein querying the persistent memory store comprises:
computing a similarity measure between the distilled problem representation and each of a plurality of indexed reasoning templates maintained in a metabuffer;
ranking the plurality of indexed reasoning templates by the computed similarity measure; and
selecting one or more reasoning templates from the plurality of indexed reasoning templates having similarity measures exceeding a predetermined relevance threshold.

### Claim 9 (depends on 1)

The method of claim 1, wherein adapting each relevant reasoning template comprises:
identifying at least one placeholder variable and at least one abstract procedural step within the relevant reasoning template;
mapping each identified placeholder variable to a corresponding parameter derived from the distilled problem representation; and
substituting each corresponding parameter into the relevant reasoning template to produce the instantiated template.

### Claim 10 (depends on 1)

The method of claim 1, wherein updating the persistent memory store further comprises applying four complementary memory management operations:
duplicate detection by comparing incoming reasoning patterns to existing entries in the persistent memory store to filter redundant patterns;
quality verification by validating stored reasoning templates;
feedback-driven promotion and demotion by adjusting relevance scores of stored reasoning templates based on observed success or failure of the templates in producing correct solutions; and
relevance re-ranking by reordering the retrieval index based on updated relevance scores to prioritize the most effective templates for future retrieval.

### Claim 11 (depends on 10)

The method of claim 10, wherein the memory management operations further comprise at least one of:
garbage collection for removing obsolete reasoning templates that have not been retrieved or have been demoted below a minimum relevance threshold over a predetermined number of problem-solving cycles;
caching frequently retrieved reasoning templates for reduced retrieval latency;
and
context generation for enriching stored reasoning templates with metadata derived from problem-solving outcomes.

### Claim 12 (depends on 1)

The method of claim 1, further comprising, prior to receiving the input problem, a cold-start initialization phase comprising:
obtaining a set of at least one of one or more representative problems or one or more problem descriptions from a target problem domain;
processing, by the dream generator module, each representative problem of the one or more representative problems through each of the plurality of parameterized prompt templates to generate an initial corpus of synthetic reasoning patterns spanning all of the plurality of functionally distinct reasoning pattern categories; and
storing the initial corpus in the persistent memory store;
wherein the persistent memory store comprises reasoning scaffolding prior to receipt of the input problem.

### Claim 13 (depends on 12)

The method of claim 12, wherein the cold-start initialization phase further comprises:
ingesting one or more past problem-solving trajectories, each past problem-solving trajectory comprising:
a problem description;
intermediate reasoning steps; and
a final solution; and
processing, by the dream generator module, each past problem-solving trajectory to extract and generalize reasoning patterns therefrom for inclusion in the initial corpus.

### Claim 14 (depends on 1)

The method of claim 1, wherein:
the method is performed without modifying weights, fine-tuning procedures, or internal architecture of the LLM agent; and
the LLM agent is interfaced exclusively through the assembled structured prompt as input and the solution as output, thereby enabling the LLM agent to be replaced with a different LLM agent without loss of accumulated reasoning templates in the persistent memory store.

### Claim 15 (depends on 1)

The method of claim 1, wherein the generating step is performed selectively, such that the dream generator module generates new synthetic reasoning patterns only when the retrieval module determines that the persistent memory store contains fewer than a threshold number of reasoning templates having similarity measures exceeding a predetermined relevance threshold for the distilled problem representation.

### Claim 16 (depends on 1)

The method of claim 1, wherein:
the structured prompt provides cognitive scaffolding that reduces the number of tokens consumed by the LLM agent in generating the solution and reduces the wall-clock time required to produce the solution, as compared to processing the input problem without the instantiated templates and the generated synthetic reasoning patterns; and
the reduction in tokens consumed and wall-clock time is greater for input problems having a complexity level exceeding a predetermined difficulty threshold than for input problems having a complexity level below the predetermined difficulty threshold.

### Claim 17 (depends on 2)

The method of claim 2, wherein the combination of all six reasoning pattern categories produces an improvement in solution accuracy that exceeds a sum of individual accuracy improvements attributable to each reasoning pattern category when applied independently.

### Claim 18

A system for augmenting the reasoning capabilities of a large language model (LLM) agent through memory-augmented reasoning enhanced by synthetic reasoning patterns, the system comprising:
one or more processors;
a non-transitory computer-readable storage medium storing instructions that, when executed by the one or more processors, cause the system to implement:
a problem distiller module configured to receive an input problem and extract therefrom a distilled problem representation;
a persistent memory store configured to store a plurality of reasoning templates organized into a plurality of functionally distinct reasoning pattern categories, the persistent memory store maintaining a memory state that persists across successive problem-solving sessions;
a MetaBuffer module configured to maintain an indexed collection of reasoning templates stored in the persistent memory store and to retrieve, based on a similarity measure computed against the distilled problem representation, one or more relevant reasoning templates;
a template instantiation module configured to adapt each retrieved reasoning template to the specific context of the input problem by binding abstract strategy descriptions and generalized procedural steps to concrete parameters derived from the distilled problem representation, producing one or more instantiated templates;
a dream generator module configured to generate one or more synthetic reasoning patterns by processing the distilled problem representation through one or more parameterized domain-agnostic prompt templates, each synthetic reasoning pattern encoding a structured problem-solving strategy belonging to one of the plurality of functionally distinct reasoning pattern categories, the reasoning pattern categories comprising at least one of: Strategy Dreams, Meta Pattern Dreams, Error Avoiding Dreams, Hard Problem Dreams, Adaptive Dreams, and Negative Examples;
a prompt assembly module configured to compose a structured prompt;
a reasoning engine comprising the LLM agent, the reasoning engine configured to process the assembled structured prompt and generate a solution to the input problem;
a quality evaluator module configured to evaluate the solution and produce a quality score; and
a memory updater module configured to update the persistent memory store from a current memory state to an updated memory state based on the quality score, the distilled problem representation, the instantiated templates, the generated synthetic reasoning patterns, and the solution, the memory updater module further comprising:
a duplicate detection component configured to filter redundant reasoning patterns against existing entries;
a quality verification component configured to validate factual accuracy and usefulness of stored templates;
a feedback loop component configured to adjust relevance scores of stored templates based on observed solution outcomes; and
a relevance ranking component configured to reorder the retrieval index to prioritize effective templates for future retrieval;
whereby the system progressively improves its reasoning capabilities over successive problem-solving interactions through continuous refinement of the persistent memory store.

### Claim 19 (depends on 18)

The system of claim 18, wherein the dream generator module is further configured to perform cold-start initialization by:
processing a set of representative problems from a target domain through each of the parameterized prompt templates to generate an initial corpus of synthetic reasoning patterns spanning all reasoning pattern categories; and
storing the initial corpus in the persistent memory store prior to any live problem-solving interaction, whereby the persistent memory store contains reasoning scaffolding from a first interaction.

### Claim 20 (depends on 18)

The system of claim 18, wherein the system operates as a model-agnostic middleware layer, and wherein the reasoning engine interfaces with the LLM agent solely through the assembled structured prompt as input and the solution as output, requiring no modification to weights, fine-tuning procedures, or internal architecture of the LLM agent, and wherein the LLM agent is replaceable with a different LLM agent while the persistent memory store and all stored reasoning templates are preserved.

### Claim 21 (depends on 18)

The system of claim 18, wherein the dream generator module is configured to operate in two modes:
a proactive mode, in which synthetic reasoning patterns of at least one of the Strategy Dreams, Meta Pattern Dreams, Error Avoiding Dreams, Hard Problem Dreams, and Negative Examples categories are generated in advance and stored persistently in the persistent memory store; and
a reactive mode, in which synthetic reasoning patterns of the Adaptive Dreams category are generated at inference time in response to specific characteristics detected in the distilled problem representation of a current input problem.

### Claim 22 (depends on 18)

The system of claim 18, wherein the distilled problem representation comprises identified features, constraints, and structural characteristics of the input problem.

### Claim 23 (depends on 18)

The system of claim 18, wherein the structured prompt comprises a rules block specifying domain constraints, a problem block containing the distilled problem representation, and a dreams block containing the instantiated templates and the generated synthetic reasoning patterns.

### Claim 24 (depends on 18)

The system of claim 18, wherein the problem distiller module comprises:
a feature extraction unit configured to identify salient numerical, structural, or semantic features of the input problem;
a constraint identifier configured to enumerate explicit and implicit constraints governing valid solutions to the input problem; and
a structure analyzer configured to characterize a problem type and assign a complexity level to the input problem, the complexity level being used by the dream generator module to selectively activate generation of Hard Problem Dreams when the assigned complexity level exceeds a predetermined difficulty threshold.
