References

Aizk combines published memory mechanisms, open source infrastructure, product comparisons, and original work. This page records which is which. A citation here does not mean that Aizk copied an implementation or that another project endorses Aizk.

How to read the map

Label	Meaning
adopted	the shipped mechanism follows the cited design closely
adapted	the source supplied the idea and Aizk changed its shape
compared	the source helped define product boundaries but supplied no implementation
workflow	the source influenced how the code was changed or checked
original	the mechanism was designed for Aizk and is not claimed from the cited systems

Memory and retrieval features

Aizk feature	Lineage	What Aizk does	Code
temporal entity and fact graph	adopted from Zep and Graphiti	stores immutable content plus valid-time and recorded-time claims	store/models/tables, store/models/views/live_fact.py
add, update, and no-op consolidation	adapted from Mem0	resolves confident cases by rule and sends only ambiguity to the LLM	graph/consolidation.py, graph/writer.py
associative multi-hop recall	adapted from HippoRAG 2	runs personalized PageRank over only visible current facts	retrieval/lanes/graph.py
community summaries	adapted from GraphRAG and LightRAG	detects communities as a rebuildable global-evidence projection	graph/communities.py, retrieval/lanes/vector.py
recursive summary tree	adopted from RAPTOR	rolls grounded summaries into bounded higher levels	graph/raptor.py, retrieval/lanes/overview.py
reflective observations	adapted from A-MEM	derives optional observations without replacing their grounding facts	graph/insight.py
entity profiles	adapted from GAM	builds one evidence-grounded profile per entity and scope set	graph/profiles.py
raw evidence as authority	supported by Does Memory Need Graphs	keeps source chunks primary and makes every graph lane earn its cost in ablation	src/aizk/retrieval/lanes/sources.py, src/eval/plans.py
append-only corrective history	supported by APEX-MEM	closes temporal ranges rather than deleting contradicted knowledge	store/models/tables/fact.py
speaker-aware group memory	adapted from GroupMemBench and Hindsight	separates objective state from observations, opinions, experiences, and preferences	src/aizk/provenance.py, src/aizk/graph/grounding.py, src/eval/groupmem.py
forgetting-aware evaluation	adopted from Memora	measures current evidence without rewarding expired memory	src/eval/metrics.py
workflow and premise benchmark categories	planned from LongMemEval-V2	keeps these categories in evaluation until a production schema earns them	src/eval
action-memory boundary	compared with Mem2ActBench	makes no action-selection claim from a retrieval-only score	docs/benchmarks.md
dense and lexical fusion	adapted from Reciprocal Rank Fusion	fuses typed source ranks inside one SQL recall program	retrieval/recall/program.py, retrieval/lanes/sources.py
merit ordering and maximal recall	original measured Aizk result	keeps every lane available and lets one cross-encoder rank candidates together	retrieval/recall/orchestrator.py, retrieval/rerank/rescore.py
public evidence provenance	original Aizk interface design	collapses internal lanes into source, derived, and session evidence and attaches exact Logto scope descriptions	retrieval/models/result.py, retrieval/templates/recall.md.j2

RAPTOR supports hierarchical summaries at several abstraction levels. GraphRAG supports community summaries for broad corpus questions. HippoRAG supports graph-based associative retrieval. GAM and A-MEM support consolidated and interconnected memory representations. These papers justify trying different internal representations, but none argues that an agent should receive Aizk’s lane names. The three public provenance classes are therefore an original interface choice based on what the consumer needs to judge evidence rather than how the engine found it.

Sharing, identity, and security

The private and shared memory distinction is informed by Collaborative Memory. Aizk changes the paper’s policy graph into one PostgreSQL-native scope lattice. Every row carries a sorted nonempty set of scope UUIDs and a reader must stand in every member. Forced row level security applies that rule to documents, chunks, graph claims, profiles, communities, and working memory.

The intersection model, full-authority reads, explicit write destinations, and source-preserving share operation are original Aizk design. Logto remains authoritative for users, organizations, roles, and public organization metadata. Aizk derives stable IDs from verified claims and stores no identity or membership mirror.

Concern	Source	Aizk boundary
identity and organization authority	Logto	OIDC discovery, signed tokens, current organization roles, and no local identity tables
OAuth protected MCP	FastMCP	dynamic client registration and an OIDC proxy backed by persistent encrypted state
database authorization	PostgreSQL row security and the house rlsalchemy package	forced policies on both content and scoped claims
multi-user memory model	Collaborative Memory	private, organization, and intersection scopes with immutable capture provenance

Infrastructure and model lanes

Responsibility	Project	Use in Aizk
relational, temporal, lexical, vector, and policy execution	PostgreSQL	one durable state engine and the preferred place for filtering, ranking, hashing, comparison, and temporal logic
vector index	VectorChord and pgvector	low-memory production vector search with a portable fallback
ORM and validation	SQLModel, SQLAlchemy, and Pydantic	typed models, PostgreSQL statements, and wire contracts
durable jobs	PgQueuer	graph projection and scheduled maintenance without a bespoke workflow ledger
document conversion	Docling and Docling Serve	private conversion of already accepted bytes into native JSON and normalized Markdown
immutable object bytes	SeaweedFS and obstore	private S3-compatible storage behind opaque keys and a bounded typed client
malware scanning	ClamAV	fail-closed streaming scan before object persistence over a private service connection
log collection	Grafana Alloy	discover Compose containers and forward labeled Docker logs without a per-service logging driver
log storage and inspection	Loki and Grafana	one bounded operational log store and a loopback-only inspection interface
log event vocabulary	OpenTelemetry Logs Data Model	structured AIZK application events while PostgreSQL remains the durable usage authority
MCP transport and OAuth	FastMCP	the four public tools and the Logto OIDC proxy
browser application	SvelteKit and @logto/sveltekit	optional server-rendered web interface for Logto login, scoped statistics, recall, artifacts, and organization management over the browser JSON API
model serving	vLLM and structured outputs	replaceable OpenAI-compatible endpoints with compact XGrammar-constrained extraction
typed LLM calls and judging	Pydantic AI and Pydantic Evals	schema-constrained extraction and isolated evaluation
chunking	Chonkie	bounded prose and source windows
fast entity gate	GLiNER2 and GLiNER2 large	a cheap GPU gate and an experimental extractor, not the production graph authority
production embedding	Qwen3-VL-Embedding-2B	text and image vectors through a generic client
production reranking	Qwen3-Reranker-4B	cross-encoder merit ordering across all retrieval lanes
production extraction	Gemma 4 12B	grounded graph extraction through the generic LLM client
profiling	the house mainboard package	stage timing, memory sizing, and production bottleneck evidence
environments and remote operation	the house chefe and lote packages	reproducible tasks and Crimson deployment
typed patterns	the house patos package	shared model and registry abstractions where they reduce code

Model names are deployment choices rather than domain names in the code. Embedding, reranking, gating, and extraction clients can move to another compatible provider without renaming the memory engine.

Original files and derived document data

Several mature systems separate an authoritative source from replaceable interpretation.

Reference	Useful mechanism	AIZK adaptation
Paperless-ngx usage and FAQ	preserve the original, track checksums, and index OCR or archive derivatives	one original Blob remains authoritative while searchable Markdown and structured data stay in PostgreSQL
Docling supported formats	emit normalized Markdown and a lossless structured Docling document	store both database derivatives on the exact original revision and keep the source for later reinterpretation
Unstructured document elements	normalize many formats into typed elements carrying source and coordinate metadata	retain source metadata and conversion details without adopting a second element store
ColPali	retrieve document pages through visual language representations without depending on lossy text extraction	supplement converted image text with a direct image vector while keeping Docling structure authoritative
VisRAG	retrieve page images and answer from visual document evidence	keep exact authorized files available on demand while avoiding automatic file transfer during recall
M3DocRAG	combine visual and textual evidence for long multimodal documents	retain both original and parsed representations while deferring page-level and video retrieval until they show measurable benefit

The papers agree that text extraction can lose layout, figures, tables, and other evidence. They differ in retrieval granularity and whether visual embeddings replace or complement text. AIZK keeps Docling text and structure authoritative, then adds one supplemental direct vector for an accepted image on the same document and exact artifact revision. The generic visual interface can later support video, but the current video path uses Docling’s audio transcript and metadata only. Page-level document images and video frames remain deferred until measured retrieval quality justifies their storage and serving cost. Unsupported conversion falls back to filename, size, media type, URI, conversion state, and companion text rather than making an accepted file disappear.

Product and project comparisons

These systems clarified the product surface but did not supply Aizk code.

Reference	What we learned	Where Aizk deliberately differs
Supermemory	excellent separation of getting started, concepts, operations, integrations, and benchmarks plus a clear document and memory mental model	Aizk exposes its engine source, makes raw evidence authoritative, and represents overlapping scope intersections under forced PostgreSQL RLS
Supermemory repository	a small client surface and MCP-first onboarding are valuable	Aizk keeps the database, extraction, ranking, temporal graph, and authorization implementation in the auditable repository
Tanka	shared memory should preserve high-density human intent from a large stream of automated execution	Aizk is memory infrastructure rather than a messenger or company operating system
EverOS	Tanka’s open framework separates atomic facts, episodes, profiles, and agentic use	Aizk does not adopt its proactive work layer or make profiles authoritative
SharedMemory	a short write and query API, isolated volumes, conflict checks, and multi-agent access are useful product language	Aizk uses three MCP verbs, database RLS, bi-temporal claims, and source-first retrieval rather than adopting its guard or volume model
Brainshared	historical inspiration for turning the best validated individual memory into a shared team floor	the site is no longer an available implementation reference and Aizk does not depend on it
Cognee	a PostgreSQL node and edge schema can unify graph and vector storage	Aizk was written from scratch as a smaller PostgreSQL-only engine with its own scope lattice
Mem0 and Letta	the useful boundary sits between a minimal memory API and a full agent framework	Aizk is memory infrastructure and does not own the agent loop
the authored Zettelkasten and qmd	exact files remain the human source of truth and the honest retrieval baseline	Aizk supplies scoped current-state recall, temporal history, and one prompt-ready result rather than replacing authored notes

Knowledge organization

Source	Inherited idea	AIZK adaptation
The PARA Method	Projects are finite action-oriented outcomes while Areas are ongoing responsibilities that need attention	Areas and Projects remain ontology entities rather than folders, and source tags associate knowledge without moving it
Combining Building a Second Brain and the Zettelkasten Method	PARA supplies action context while a Zettelkasten supplies atomic durable knowledge and structure notes that link supporting material	one maintained brief maps each Area or Project, while atomic supporting notes use #project: <name> and #area: <name>
the user’s historical Zettelkasten Structure Note	#project and #area identify structure notes and every Project belongs to an Area	key-value source tags name the exact entity, work for any live ontology kind, and never imply status or access

“Shared brain” is Aizk’s product description. The phrase also acknowledges the historical Brainshared product idea, but no Brainshared package or code runs inside Aizk. Here it means that several people and MCP agents can use one memory service while database policy preserves private, shared, and intersection knowledge.

Engineering workflow references

The Bun Rust rewrite report and its original PORTING.md commit influenced the refactor process rather than the runtime. The reusable ideas are a written mapping before a broad change, small trial cells, bounded ownership, independent adversarial audit, errors as a work queue, and a language-independent test suite as the final authority. Aizk does not copy Bun code and does not treat compilation as proof of behavior.

The documentation information architecture takes inspiration from Supermemory’s current docs. The text, examples, diagrams, visual system, and claims here are original to Aizk.

Rejected and deferred ideas

References also matter when the evidence says not to ship a mechanism.

Idea	Status	Reason
query-time routing	rejected	the measured classifier was only 44 percent accurate and a wrong route could remove decisive evidence
fixed lane ordering	rejected	facts-first and overview-first each failed on a different evaluation stratum
GLiNER2 as graph authority	rejected for production	the large GPU model is fast but still misses dense relations that the LLM extracts
graph-only authority	rejected	source briefs and chunks are more faithful than generated profiles and summaries
human acceptance queue	permanently rejected	AIZK has no acceptance layer because agents write sources directly and correct them when evidence changes
bespoke graph workflow ledger	rejected for now	the graph is a rebuildable projection and PgQueuer already owns durable execution state
TRACE-KG ontology induction	studied, not adopted	Aizk currently prefers a declared ontology with deterministic validation
Tanka-style proactive work layer	deferred	Aizk supplies memory and does not own the agent’s planning or action loop
Supermemory-style connector breadth	deferred	current work prioritizes correctness, access, time, and evaluation over ingestion breadth

Original Aizk contributions

The following pieces should be cited as Aizk design rather than attributed to one upstream paper.

• arbitrary nonempty scope sets with intersection visibility under forced PostgreSQL RLS
• content-addressed graph content separated from scoped bi-temporal claims
• full-authority recall with one explicit write destination
• a source-preserving share operation that creates provenance-linked copies
• one maximal recall plan whose cross-encoder orders every lane by merit
• a single prompt-ready MCP recall string produced by a token-budget prefix
• exact artifact revision resources that remain authorized by PostgreSQL and transfer no bytes during recall
• one original-only Blob model with database derivatives, metadata fallback, adaptive compression, shared physical bytes, a fail-closed scan gate, and no Redis
• one supplemental direct image vector on the authoritative converted document with exact revision provenance and no model-provider coupling
• durable actor and scope usage accounting separated from expiring operational logs
• an operational health snapshot that checks schema, policy, jobs, models, scopes, graph freshness, and a real recall in under five seconds

The Benchmarks page records what has been measured. The Comparison page separates observed results from architectural differences. Neither page turns a related paper, model, or product into an endorsement.