Code graph and memory for your coding agents, in one Rust program. For Claude Code, Codex, Cursor and any MCP client.
Third-party attribution for the Graphicle memory library
Copyright (c) 2026 KarLex AI, Inc. The memory library is licensed under the Apache License, Version 2.0 (LICENSE). Adapted third-party portions keep their upstream MIT or Apache-2.0 terms; the notices of the Apache-2.0 sources are reproduced in the repository's NOTICE and their licence texts are in licenses/. This is not a relicensing of any upstream repository, nor a claim of original authorship of its algorithms. Enterprise-licensed Onyx code is excluded.
Selected implementations and designs were translated or adapted from the pinned sources below. This is not a port of each entire application. Modified portions are identified in UPSTREAM.json. Upstream MIT and Apache-2.0 terms continue to apply to their adapted portions.
Graphify-Labs/graphify
https://github.com/Graphify-Labs/graphify/tree/20a20d30d8e7eef77675651f0199d87f913bd3e7
49 pinned syntax grammar selections, declarations, imports, calls and original source coordinates; expanded generic language analyses remain incomplete and do not claim compiler type resolution.
License and notices: licenses/Graphify-Labs__graphify__LICENSE.txt, licenses/Graphify-Labs__graphify__LICENSE-MIT.txt, licenses/Graphify-Labs__graphify__NOTICE.txt
MemTensor/MemOS
https://github.com/MemTensor/MemOS/tree/a7367d07e55db61099f7b4e2c1108bc5831a24f3
Scoped exact-reuse cache and prioritized projection readiness adaptations. The Rust cache stores lexical features and source layouts, not embeddings; no Redis queue or scheduler service port.
License and notices: licenses/MemTensor__MemOS__LICENSE.txt
getzep/graphiti
https://github.com/getzep/graphiti/tree/16cdf7045378c8d53ae01f94e2fa60d238cb0f68
RRF, explicit-clock contradiction intervals and contextual semantic task contract. Candidate ties are deterministic by ID; unknown clocks remain unknown.
The alias ladder of src/graph/canonical.rs (exact normalised names, then a Shannon-entropy gate, then MinHash Jaccard of character 3-gram shingles of at least 0.9) follows graphiti_core/utils/maintenance/dedup_helpers.py at the same commit.
Copyright 2024, Zep Software, Inc. (the copyright notice of the adapted Graphiti files, retained in src/retrieval.rs, src/temporal.rs, src/semantic.rs and src/graph/canonical.rs).
License and notices: licenses/getzep__graphiti__LICENSE.txt
mem0ai/mem0
https://github.com/mem0ai/mem0/tree/a39a802bbc93e85b820078cd3c4dbaf53af25dbe
Contextual, attributed fact extraction and update task contract; connected model supplies interpretations. Instructions to conceal actual provenance are not retained.
Later revision https://github.com/mem0ai/mem0/tree/94c3fe9f238f3dbf29c9ce98643bd71eb13077cd: rule-based entity candidates (mem0/utils/entity_extraction.py) adapted in src/entities.rs without spaCy, to name the question terms no source mentions.
License and notices: licenses/mem0ai__mem0__LICENSE.txt
microsoft/graphrag
https://github.com/microsoft/graphrag/tree/82b87bf0434fc69f857663280e2c36a1c48e52fe
Distinct-source community weights and token-bounded context for supplied reviewed reports; no automatic community report synthesis.
License and notices: licenses/microsoft__graphrag__LICENSE.txt
onyx-dot-app/onyx
https://github.com/onyx-dot-app/onyx/tree/4d0a9c9ea3bd69d4ac4a3de03db93d27dec90307
Token-bounded chunk accumulation adapted to preserve exact UTF-8 original source bytes. Enterprise code and connector catalogue excluded.
Later revision https://github.com/onyx-dot-app/onyx/tree/b2186a2dba8123ab597e3d28043b6e64619bf135: multi-query fusion. Case-insensitive query deduplication that adds the weights of repeats (deduplicate_queries in backend/onyx/tools/tool_implementations/search/search_tool.py) and one weighted reciprocal-rank-fusion ranking per query (weighted_reciprocal_rank_fusion in search_utils.py of that directory), adapted in src/retrieval.rs (opt-in query_variants; each variant fuses as a weighted lexical_q{i} arm, and a caller's dense ranking for it as semantic_q{i}). The variant kinds follow RAG-Fusion (plain rephrasings), HyDE (Gao et al., 2022) and query2doc (Wang et al., 2023) for hypothetical answer passages; implemented independently from the papers.
License and notices: licenses/onyx-dot-app__onyx__LICENSE.txt
vectorize-io/hindsight
https://github.com/vectorize-io/hindsight/tree/680406b3dd9cca2108c7f0e204820a09b4e30906
Version-2 opaque document blocks, eight delta operations, missing-grounding detection and observation contract. Input bounds, stricter fences, transitive invalidation and approval-separated application added.
Later revision https://github.com/vectorize-io/hindsight/tree/eb021da3b2501911e4b57c82b3de1123572a200e: English period rules from hindsight_api/engine/temporal_periods.py adapted in src/temporal.rs (temporal.window).
Later revision https://github.com/vectorize-io/hindsight/tree/eb021da3b2501911e4b57c82b3de1123572a200e: resolve_section_anchor (an add_section anchor may name a section added earlier in the batch, a slug or a heading) and the wrong-section block report of resolve_block, from hindsight_api/engine/reflect/delta_ops.py, adapted in src/document.rs (document.delta).
License and notices: licenses/vectorize-io__hindsight__LICENSE.txt
Retrieval designs read 2026-09-30
These are independent implementations of published designs; no upstream code was copied. design_adaptations in UPSTREAM.json records the inspected files and their hashes.
- microsoft/Memora, https://github.com/microsoft/Memora/tree/dec3f8f2444eace7004fc084abe1be9f3d88270e: a weight per ranking in reciprocal-rank fusion (
arm_weights,rrf_kinsrc/retrieval.rs), after_merge_results_with_rrfinsrc/memora/core/memory.py. License: licenses/microsoft__Memora__LICENSE.txt - openmemind/memind, https://github.com/openmemind/memind/tree/7536460664228a7b48e41c4ed61f2a75af244b17: weighted fusion and the strong-signal rule of
evidence_strength, afterResultMerger.javaandScoringConfig.java. License: licenses/openmemind__memind__LICENSE.txt - HKUDS/LightRAG, https://github.com/HKUDS/LightRAG/tree/453dce83d6d0354a06e46c8d4029a0895c4e054b: per-origin quotas falling linearly from
maxtominwith leftover polling (origin_quota), afterlightrag/operate.pyandlightrag/utils.py. License: licenses/HKUDS__LightRAG__LICENSE.txt - FreedomIntelligence/Tiermem, https://github.com/FreedomIntelligence/Tiermem/tree/dba57bf2ce91a9920fdebaf5ffb0e6fc029fc398: weak-evidence and aggregate-question features of the
evidence_strengthreport, aftersrc/linked_view/router.py. License: licenses/FreedomIntelligence__Tiermem__LICENSE.txt - xiaowu0162/LongMemEval, https://github.com/xiaowu0162/LongMemEval/tree/9e0b455f4ef0e2ab8f2e582289761153549043fc: a further ranking over the user's own turns (opt-in
speakerinsrc/retrieval.rs; the"auto"choice of the assistant's or the user's side from the question is Graphicle's own), after the user-only retrieval keys ofprocess_item_flat_indexinsrc/retrieval/run_retrieval.py; the opt-inkey_expansion: "merge", which adds a row'sabstractionsto its lexical text, after the key = value + fact join ofresolve_expansioninsrc/retrieval/index_expansion_utils.py. License: licenses/xiaowu0162__LongMemEval__LICENSE.txt - SmartSearch (arXiv 2603.15599; no code published): term-focused window trimming (
pack_focusinsrc/packing.rs), from the paper's description. The paper is licensed CC BY 4.0 (Derehag, Calva and Ghiurau). - M. F. Porter, "An algorithm for suffix stripping", Program 14(3), 1980, and the reference implementation's two documented departures, https://tartarus.org/martin/PorterStemmer/: English suffix stemming of BM25 terms (opt-in
stem,src/stem.rs), as in Lucene's and Anserini's English BM25 baselines.
Rust dependencies
Cargo.lock pins the linked and build-time dependencies. The licence texts of the crates linked into a release binary are in the THIRD_PARTY_LICENSES.html of the program that ships this library (../graphicle-graph/licenses/THIRD_PARTY_LICENSES.html for the graph engine, ../graphicle/THIRD_PARTY_LICENSES.html for the graphicle program), generated from Cargo.lock by cargo about. Grammar-specific licenses and pinned revisions are in GRAMMAR_NOTICES.json, and each grammar project's licence text is in licenses/grammars/. Licence texts recovered from the source repositories of crates whose packages carry none are in licenses/upstream-crates/ (licenses/CRATE_SOURCE_NOTICES.json). Declared-license fallbacks are identified in licenses/DECLARED_CRATE_NOTICES.json; they are not represented as recovered upstream copyright notices.