Files

kubernetes/apps/ai/memini/app/helmrelease.yaml

1 ---

2 # yaml-language-server: $schema=https://k8s-schemas.home-operations.com/helm.toolkit.fluxcd.io/helmrelease_v2.json

3 apiVersion: helm.toolkit.fluxcd.io/v2

4 kind: HelmRelease

5 metadata:

6 name: memini

7 spec:

8 chartRef:

9 kind: OCIRepository

10 name: memini

11 interval: 30m

12 values:

13 controllers:

14 main:

15 type: deployment

16 replicas: 1

17 annotations:

18 reloader.stakater.com/auto: "true"

19 pod:

20 securityContext:

21 fsGroup: 1000

22 fsGroupChangePolicy: OnRootMismatch

23 runAsGroup: 1000

24 runAsNonRoot: true

25 runAsUser: 1000

26 containers:

27 main:

28 env:

29 MEMINI_API_KEY:

30 valueFrom:

31 secretKeyRef:

32 name: memini-secret

33 key: MEMINI_API_KEY

34 MEMINI_BACKEND: sqlite

35 MEMINI_DEFAULT_NAMESPACE: default

              MEMINI_CLIENT_DEFAULTS: '{"inject_briefing_pinned":10,"inject_briefing_max_tok":3000,"recall_limit":5,"inject_recall_max_tok":400}'

37 MEMINI_DEDUP_SIMILARITY: "0.80"

38 MEMINI_DEDUP_TIERS: semantic,procedural

39 MEMINI_DEDUP_LLM_MERGE: "true"

40 MEMINI_SHORT_TERM_CAP: "300"

41 MEMINI_DEMOTE_AFTER: 1440h

42 MEMINI_EMBED_BASE_URL: http://bge-embedding-m3.ai.svc.cluster.local:8080/v1

43 MEMINI_EMBED_MODEL: bge-m3

44 MEMINI_EMBED_DIMS: "1024"

45 MEMINI_EMBED_MAX_CONCURRENCY: "2"

46 MEMINI_LOG_FORMAT: json

47 MEMINI_LLM_API: openai

48 MEMINI_LLM_BASE_URL: http://litellm.ai.svc.cluster.local:4000/v1

49 MEMINI_LLM_MODEL: opencode-go/glm-5.2

50 MEMINI_LLM_API_KEY:

51 valueFrom:

52 secretKeyRef:

53 name: memini-secret

54 key: LITELLM_API_KEY

55 MEMINI_TOMBSTONE_TTL: 720h

56 MEMINI_WRITE_EMBED_TIMEOUT: "30s"

57 MEMINI_RECALL_EMBED_TIMEOUT: "10s"

58 MEMINI_RERANK: http://bge-reranker-v2-m3.ai.svc.cluster.local:8080/v1

59 MEMINI_RERANK_MODEL: bge-reranker-v2-m3

60 MEMINI_RERANK_POOL: "10"

61 MEMINI_RERANK_MAX_DOC_CHARS: "1024"

62 MEMINI_RERANK_MAX_BATCH_CHARS: "48000"

63 MEMINI_RERANK_TIMEOUT: "10s"

64 MEMINI_RERANK_MIN_SCORE: "0"

65 resources:

66 requests:

67 cpu: 25m

68 memory: 128Mi

69 limits:

70 memory: 512Mi

71 securityContext:

72 allowPrivilegeEscalation: false

73 readOnlyRootFilesystem: true

74 capabilities:

75 drop:

76 - ALL

77 fsck:

78 enabled: true

79 cronjob:

80 schedule: "0 5 * * *"

81 containers:

82 fsck:

83 env:

84 MEMINI_API_KEY:

85 valueFrom:

86 secretKeyRef:

87 name: memini-secret

88 key: MEMINI_API_KEY

89 persistence:

90 data:

91 existingClaim: memini

92 route:

93 ui:

94 enabled: true

95 annotations:

96 gatus.home-operations.com/enabled: "false"

97 parentRefs:

98 - name: envoy-internal

99 namespace: network

100 hostnames:

101 - memini.janpuc.com

102 serviceMonitor:

103 main:

104 enabled: true

105 grafanaDashboards:

106 enabled: true

107 folder: ai

dedup LLM merge overwrites the survivor, not the tombstones

gotcha ·ok L37

LLM merge folds each near-duplicate cluster into one comprehensive memory rather than arbitrarily keeping one member's text. It needs the LLM, which only started working on 2026-07-29.

Show the rest

Understand the asymmetry before touching MEMINI_DEDUP_SIMILARITY. Tombstoned duplicates stay recoverable — SetSuperseded marks them and TOMBSTONE_TTL keeps them 30 days. The SURVIVOR's content does not: it is overwritten via Upsert with no in-store history, so only a PVC restore brings the original back. A false-positive cluster is therefore far more expensive with merge on than it was with merge off, and 0.80 is the aggressive end of what is safe at this setting.

2026-08-0401KZ6M12JQYB6SNVTPSYNDXD6N

MEMINI_EMBED_MODEL value pinned across memini and the embedder

invariant ·ok L43

Stays "bge-m3" even though the InferenceService next to it is named bge-embedding-m3. memini persists this string in its store and refuses to boot if it ever disagrees, so renaming it to match the resource would force a full re-embed of every memory for no benefit — llama.cpp still serves the identical weights under --alias bge-m3. The matching invariant is recorded on kubernetes/apps/ai/llmkube/models/bge-embedding-m3.yaml; the two must be changed together or not at all.

2026-08-0401KZ6M15GQN4T4EE6KMMPGX8R8

rerank pool and doc cap sized to fit inside MEMINI_RERANK_TIMEOUT

why ·ok L60

Sized from measurements on this iGPU: roughly 130 ms/doc at 350 chars and 335 ms/doc at 1024. Latency is linear in POOL x doc length, so a pool of 20 at the 2048 cap would take about 14s, blow MEMINI_RERANK_TIMEOUT, and silently drop every recall back to composite ordering — the failure is invisible, which is what makes it worth pinning. 10 x 1024 holds the worst case near 3.4s (typically ~1.3s) while still giving three times the recall limit of rescue depth, which is the whole point of having a pool.

2026-08-0401KZ6M19R9NHFS90NSTPWQ5V8E

rerank batch chars raised so a 20-doc pool ships in one call

why ·ok L62

Raised far above the 6000 default so the entire pool goes to /rerank in a single call. At 6000 a 20-doc pool shards into roughly 7 SERIAL rerank calls, which blows MEMINI_RERANK_TIMEOUT.

2026-08-0401KZ6M1BT1B1BP4XAKS3R973GB

positive RERANK_MIN_SCORE blindfolds personal recall

anti-pattern ·ok L64

Disabled deliberately. Do not re-enable without re-measuring.

0.1 was tried on 2026-07-30 to make recall return empty on nonsense queries, and it did exactly that. But bge-reranker emits unbounded logits and scores generically-phrased questions about the user NEGATIVE: "user profile" -0.35, "who is the user" -4.10, against "tell me about Janek" +0.34. Any positive floor therefore blindfolds precisely the personal/jan recalls that matter most — Open WebUI answered "I don't have any personal memories about this user" while 39 of them sat in the store.

Show the rest

A negative floor is not an escape hatch: config.validate() refuses MEMINI_RERANK_MIN_SCORE < 0 at boot, so this knob is on/off only. Disabling it costs nothing in ordering, which is the reranker's real value; relevance is still gated by inject_recall_min_score on the injection path.

2026-08-0401KZ6M1FA18P9JBVV0V4V0K39R