AuraSDK (repository name aura-memory) is a local cognitive memory runtime that can run alongside any frozen model, providing AI agents with persistent memory and an auditable cognitive layer. Its core idea is: the model itself remains unchanged, and agent performance is enhanced by locally accumulating structured experience, with no fine-tuning, no embedding vectors, no cloud training, and no API keys required. The project is written in pure Rust, offline-first, and patent pending.
Purpose and Positioning Aura is aimed at agent scenarios that require cross-session semantic memory, transforming prompt-only, forgetful agents into auditable, memory-aware systems. It emphasizes determinism and local recall, and does not require embeddings or LLM calls. The official documentation explicitly states: Aura is a free and local third path—different from expensive fine-tuning, and different from RAG, which requires embeddings and vector databases.
Main Capabilities Memory and Recall: Provides persistent recall, temporal versioning, and provenance tracking. recall() returns source context by default, with first-hand memory taking priority; web, document, and tool content is fenced off and cited as data, preventing instructions in external text from influencing the model. Security Configuration: Supports security="strict", set_security_profile(), and security_report(). Memories written via MCP are marked as inferred rather than the user's original words, preventing a single untrusted source from repeatedly asserting claims and crowding out first-hand facts. Multi-layer Cognitive Pipeline: Record → Belief (±5%) → Concept (±4%) → Causal (±3%) → Policy (±2%); each layer can be configured individually or enabled all at once. Higher layers also provide read-only advisory outputs, such as stable concept abstractions, learned causal patterns, and policy hints (Prefer / Avoid / Warn), which do not automatically affect behavior. Explainability and Correction: Provides explain_recall(), explain_record(), provenance_chain(), explainability_bundle(); supports targeted retraction/deprecation APIs, a persistent correction log, a correction review queue, and advisory corrections (not automatically applied). Bounded Autonomous Plasticity: capture_experience(), ingest_experience_batch(), maintenance-phase integration, anti-hallucination guards, plasticity risk scoring, andClear/freeze control. Contradictions and reflection: salience markers (mark_record_salience(), etc.), reflection summaries (get_reflection_summaries(), etc.), belief instability summaries, contradiction clusters and contradiction review queues, and marking unresolved evidence in recall explanations. Evidence and decision audit graph: a deterministic read-only model preserving the complete chain from evidence to verification results, covering directed relations such as supports, refutes, contradicts, supersedes, derived_from, recalled_for, used_evidence, used_by, caused, produced, verified_by, etc.; the state is append-only and bitemporal, and historical reconstruction uses both record time and business-time validity periods. Observational outcome receipts: can record candidate record IDs, selected IDs, and classified external outcomes, without storing prompts, answers, or memory payloads; supports schema-v1 and schema-v2 (including fields such as logging policy and selected-set probabilities), is idempotent on the key (namespace, task_id, attempt_id), rejects conflicting retries, and receipts carry an integrity digest and do not affect recall, strength, retention, merging, or forgetting. Immutable evidence lineage: binds extracted claims to the exact bytes of immutable source revisions, with VerificationStatus and AnswerPermission as independent gates; high confidence cannot override source hash changes, superseded claims, or blocked references. Context capsule: can request a deterministic hot-context projection, preferentially presenting refutation scars, pending evidence debts, active goals, contradictions, outcomes, decisions, and persistent domain/identity records, returning an estimated token count, number of omitted records, selection rationale, and a stable content hash; blocked records and superseded versions outside their validity period will not appear. Supports deterministic loss-aware packing, removing exact duplicate sentences and safe discourse prefixes, without calling an LLM and without changing stored records; active goals and non-text payloads are excluded from compression. Citation-locked retrieval turn: for answers that must be auditable, tracks the evidence actually opened during a retrieval turn, requiring every atomic answer claim to be covered by an opened and citable source; changed, superseded, disputed, background-only, blocked, unopened, or partially supported evidence cannot authorize an answer.
Usage and Integration Aura can run standalone or be integrated into multiple AI tools via MCP. The Rust features mcp-http and mcp-bridge provide a token-protected loopback MCP server and an aura-bridge stdio bridge, enabling multiple clients to share the same storage; the Aura for Windows desktop application is based on this. On the Python side, HTTP service support can be installed via pip install "aura-memory[http]", python -m aura serve refuses to listen beyond localhost when no API key is configured; python -m aura capture is a Claude Code hook used to preserve the user's original words as first-hand memory.
Use cases Suitable for developers and teams who want to build a local, auditable, explainable memory system for agents, especially for scenarios requiring cross-session persistent memory, source tracing, contradiction governance, multi-tool shared memory, and evidence auditing of answers.
Limitations and notes No unmeasured speed, cost, or feature ranking comparisons are officially published; recall quality and speed are measured separately, and all data are observations rather than guarantees, affected by corpus size, query shape, cache state, enabled features, and background load. Weak records in cold logs will not automatically be reactivated through text queries. Features such as the evidence graph and receipts are observational or advisory and do not change memory behavior. Retrieval turns are transient and read-only, and do not activate or rewrite memory. Some advanced capabilities (such as adjusting ranking or retention policies based on receipts) remain limited by independent time-held-out validation.
